text stringlengths 1.23k 293k | tokens float64 290 66.5k | created stringdate 1-01-01 00:00:00 2024-12-01 00:00:00 | fields listlengths 1 6 |
|---|---|---|---|
Edge-Based Efficient Search over Encrypted Data Mobile Cloud Storage
Smart sensor-equipped mobile devices sense, collect, and process data generated by the edge network to achieve intelligent control, but such mobile devices usually have limited storage and computing resources. Mobile cloud storage provides a promising solution owing to its rich storage resources, great accessibility, and low cost. But it also brings a risk of information leakage. The encryption of sensitive data is the basic step to resist the risk. However, deploying a high complexity encryption and decryption algorithm on mobile devices will greatly increase the burden of terminal operation and the difficulty to implement the necessary privacy protection algorithm. In this paper, we propose ENSURE (EfficieNt and SecURE), an efficient and secure encrypted search architecture over mobile cloud storage. ENSURE is inspired by edge computing. It allows mobile devices to offload the computation intensive task onto the edge server to achieve a high efficiency. Besides, to protect data security, it reduces the information acquisition of untrusted cloud by hiding the relevance between query keyword and search results from the cloud. Experiments on a real data set show that ENSURE reduces the computation time by 15% to 49% and saves the energy consumption by 38% to 69% per query.
Introduction
In the era of the Internet-of-things, the explosive growth of smart sensor-equipped mobile devices enables the user to sense, collect, and process more data for intelligent control [1]. The sensor data increases rapidly and touches all aspects in human life. How to store the sensor data efficiently and securely is essential to make use of them to achieve intelligent control [2,3].
Mobile cloud storage refers to the access on mobile devices to cloud storage [4,5]. It enhances mobile experience that was previously impossible on resource-constrained mobile devices. Data owners are motivated to store their sensitive sensor data in cloud storage due to its great accessibility and low cost. However, mobile cloud storage brings not only convenience, but serious network security problems [6][7][8]. For example, in the environment of a smart grid, some sensor nodes (e.g., smart meters, power grid health monitoring sensors) are deployed in every family to sense and collect user fine-grained energy consumption data and send the data to the cloud storage which the power grid operators run [9]. It enables the power grid operators to effectively manage the demand and supply of electricity, which is significant in withstanding large-scale blackouts caused by insufficient power supply during peak times. However, it also benefits attackers to know about the resident power consumption mode, and increases the possibility of a crime, like burglary. Thus, it is the search results, which reduces the information acquisition of the cloud and prevents an untrusted and uncontrollable cloud from eavesdropping on the user's data.
The paper is organized as follows: Section 2 introduces the related works. Section 3 presents the traditional encrypted search architecture over cloud data. Section 4 describes the design of ENSURE. The evaluation results are given in Section 5. Finally, we summarize our conclusion in Section 6.
Related Work
Mobile cloud storage faces various security threats. Firstly, the cloud provider may be untrusted and want to profit from user information [23,24]. Secondly, several users share the same physical infrastructure in cloud storage, thus malicious users can obtain other user information through attacks such as unauthorized access, reverse control, and memory leaks [25]. Lots of privacy protection algorithms are proposed to resist the security risks caused by the untrusted cloud.
Encrypted search schemes are one of the privacy protection algorithms. They can be categorized into two main classes: boolean keyword search and ranked keyword search. The former selects files only based on whether the keyword appears in the content of the file and is not concerned about any relevance of the files in the search result (e.g., [12][13][14]26]). The latter records the relevance scores to compare the relevance of files to the searched keyword and then replies with the top-k relevant files as a response (e.g., [11,15,16]). It greatly improves the efficiency of extracting useful information and accuracy. Therefore, it has been widely used by cloud storage and has attracted many researchers to develop it.
Most of the latest studies about ranked keyword search focus on the cloud storage scenario. In [11], Swaminathan et al. developed a framework for a confidentially-preserving rank-ordered search. In this scheme, the computation task of the relevance scores is assigned to the client side, which increases the client's workload as a sacrifice to ensure the data security. In [10], Zerr et al. proposed the concept of r-confidentiality as the degree of information leaked from an index, and proposed a system that allows for tunable index confidentiality and efficiency. But it only allows the client to decrypt the posting list and perform a top-k relevant search. This kind of research all assigns a heavy workload to the client, so it is not suitable for mobile cloud storage which uses resource-constrained mobile devices to access the cloud service.
Considering the resource constraints in a mobile device, Miettinen and Nurminen in [27] pointed out that offloading some computing intensive tasks onto a cloud could be an effective way of dealing with this issue. Wang et al. in [26] presented a secure ranked keyword search over encrypted cloud data which used one-to-many mapping Order Preserving Encryption to encrypt the index of the file set. This design allows an efficient server-side ranking which reduces the client-side workload. Li et al. in [14] used cloud computing to improve the encrypted data search performance. Bowers et al. [28] proposed a distributed cryptographic system where the safety and retrieval of stored files are proven by a set of cloud servers. Wang et al. [15] introduced a secure ranked keyword search over encrypted cloud data. However, information leakage relating to the relevance between keywords and documents exists in these schemes, which could result in untrusted cloud providers obtaining the major term of stored files.
In the era of the Internet of Things, edge computing is rapidly emerging because edge servers build a bridge between resource-constrained mobile devices and the cloud. Since applications can access infrastructure and application services provided on-premises [29], and edge servers and mobile devices are usually in the same LAN, edge severs would be conductive for mobile devices to implement the necessary privacy protection algorithm and protocols efficiently with their relatively adequate resources when data owners want to outsource massive local data from mobile devices to the cloud. Wadood Abdul et al. [30] proposed implementing visual cryptography and zero-watermarking algorithms on edge servers to encrypt the face images that are to be uploaded to the cloud. This prevents the untrusted cloud from obtaining and abusing user biometrix content, and also ensures the image quality, which means that the result of face recognition is not affected by image encryption. Maher Jridi et al. [31] also offloaded the image compression and encryption tasks to the digital gateway on the edge of network. Manisha Jindal et al. [32] used a trustworthy edge server to implement a secure forward encryption algorithm to prevent data from being acquired by unauthenticated users and untrusted service providers. Therefore, edge servers can be a key component in the secure data processing framework for mobile cloud storage and resemble the cloud trusted domain which is introduced in [33].
Therefore, we regard the ranked keyword search scheme as a basis to establish an efficient and secure encrypted search architecture over mobile cloud storage with the help of edge computing.
Traditional Encrypted Search over Cloud Data
The traditional encrypted search architecture over cloud data is illustrated in Figure 1. Data owners and data users have different procedures to be completed. Index generation and file encryption should be processed by data owners, while data users should accomplish encrypted file search and retrieval.
Index Generation and File Encryption
When the data user intends to perform a keyword search on stored files, downloading all these files from the cloud and decrypting them to find out the top-k relevant files on the client side wastes great time and energy. The solution can be storing an extra file index in the cloud for searching. The highly effective index mechanism is the base of the highly effective search. So, the data owner first extracts distinct keywords from stored files and then builds a secure file index, which is usually an inverted file [34] consisting of a sequence of posting lists. Specifically, every keyword needs to be encrypted and hashed to fix its entry in the index, and the related posting list also needs to be encrypted against unauthorized access. The file encryption procedure is simple for mobile devices, just encrypted in a way that is consulted with the data user in advance. Lastly, both the file index and the encrypted files are sent to the cloud.
Encrypted File Search and Retrieval
In the process of search and retrieval, the cloud helps the user to find the top-k relevant files according to the keyword which the user submits. It is noteworthy that only authorized users who are capable of generating the trapdoor are entitled to search the encrypted files. The trapdoor is used to search for the intended keyword in the cloud. The process is divided into five steps. For better understanding this process, we illustrate the five steps in Figure 2, and present the related computational components for these steps.
1.
Keywords Processing: When users submit the keyword, the client first processes the keyword to generate the trapdoor of the keyword. Then, the client sends a search request (the trapdoor of the keyword) to the cloud server.
2.
Index Searching: On receiving the search request, the cloud uses the trapdoor to gain entry to the file index. Then, the posting list related to the keyword is sent back to the data user.
3.
Calculation&Rank: The data user decrypts the posting list and calculates the relevance scores to find the top-k relevant files, and then sends a request to retrieve the files.
4.
File Retrieval: The cloud server finds the target files and sends them back to the data user.
5.
File Decryption: The data user decrypts the target files to recover the original data.
Challenges in Mobile Cloud Storage
Efficiency Challenges: in the traditional encrypted search scheme, data processing tasks are mainly handled by the user's device, and many of the tasks are computation-intensive, such as decryption and hash. However, the encrypted file search and retrieval procedure is not suitable for mobile devices for two reasons. It is a heavy burden to decrypt the index and calculate the relevance scores for a mobile device. The mobile device may take more time to complete the calculation step than a powerful PC, so the query response time increases significantly and user experience is degraded. From the aspects of energy consumption, this design is also inappropriate. More latency and energy consumption will be introduced by a large amount of communication between the mobile device and cloud, and at the same time, the large traffic consumption may cost a payable traffic fee. Therefore, if we directly transplant this traditional solution into mobile cloud storage, that may lead to a low-efficiency system and a poor user experience for mobile users.
Security Challenges: offloading some computation intensive tasks onto the cloud is a popular way to compensate for the drawback of the resource-constrained mobile devices and is widely used in many mobile applications [35]. But in the scenario of an encrypted search, we usually assume that the cloud is "honest-but-curious" [12], and attempts to access the underlying plaintext of users' data. Additionally, there are many attack models [17,36,37] to use the information leakage against the encrypted search scheme. A desired security scheme should minimize the information leakage. Therefore, offloading computation onto the cloud is not an appropriate way to improve the efficiency of the encrypted search scheme in consideration of the data security. Moreover, in an encrypted search scheme, the computation in the untrusted cloud should be as little as possible in order to minimize the sensitive information leakage. For example, in the traditional ranked keyword search scheme, if the cloud server is in charge of the calculation task of the relevance scores, a more practical performance may be achieved. However, the cloud may deduce the association between keywords and encrypted files, which may cause the major subject of a document learned by the cloud [17]. In that case, a great threat to user's data security may be posed.
We have made clear the challenges ahead, so we design ENSURE to achieve the following goals, and the detailed system design will be introduced in the next section.
1.
Improve the performance efficiency of traditional encrypted search method, which includes reducing the file search/retrieval time and energy consumption.
2.
Try our best to minimize the information acquisition of the curious cloud.
ENSURE System Design
To effectively address the challenges and achieve our goals, we introduce a new architecture named ENSURE. In this section, we first discuss why we have chosen edge computing to solve the problems, and then introduce the design idea of our system and the process of file search and retrieval in ENSURE. Finally, we discuss the performance efficiency of ENSURE.
Edge Computing
Edge computing refers to the enabling technology which allows computation and service to be hosted at the edge of the network, and plays the role of the middle layer between the data source and cloud. There are various types of resources at the edge of the network. Typically, the mobile terminal whose energy and compute capability are limited (e.g., sensors, mobile phones, and tablets) is regarded as an edge node, while the device which has a constant energy supply and continuous network connectivity (e.g., home PCs, laptops) is regarded as an edge server. Note that edge node and edge server are relative concepts. There is no clear boundary between the two.
It brings new opportunities to the scenario with multiple types of sensors. This is because the edge sever could pull some tasks up from sensor-equipped mobile devices and some appropriate services can also be pushed from the cloud toward the edge. For example, in the smart home system, lots of wireless sensor-equipped mobile devices are deployed to sense, collect, and process data to achieve intelligent control, such as the temperature sensor and pyroelectric infrared sensor. If the data is leaked, the attacker could judge whether the owner is at home, or get other useful information. But such mobile devices are constrained in computation and storage resources, resulting in the lack of data protection. There are numerous similar scenes where multiple sensor-equipped mobile devices move in a certain region, such as hospitals and supermarkets. Edge computing suggests users choose one of the available, controllable, trustworthy, and relative resource-rich local devices to be the edge server to settle sensor-equipped mobile devices' annoyance. Additionally, the edge server would provide the interface for various sensors to gather and process sensed data more efficiently. With this approach, sensor-equipped mobile devices improve the performance and prolong the time of use at the expense of the energy of the edge server. Edge servers usually have more energy reserve and are easier to provide with an energy supply, and mobile devices with less energy reserve need to be recharged frequently. What is worse, users may have to recharge several powerless mobile devices simultaneously, which really reduces the user experience. Meanwhile, the extra overhead is generated by the interaction between edge servers and mobile devices, but it is quite small relative to the benefit that the edge server brings.
The Basic Idea of ENSURE
Inspired by the characteristic of edge computing, ENSURE leverages edge servers to achieve an efficient and secure ranked keyword encrypted search. It is based on the following assumptions:
1.
Sensor-equipped mobile devices move in some certain region.
2.
One of the available, controllable, trustworthy, and relative resource-rich local devices in that region plays the role of edge server.
3.
The data transmission between mobile devices and the edge server is safe.
The basic idea of ENSURE consists of two parts. The first part aims to improve the efficiency of the encrypted search. We propose to offload the workload of Calculation&Rank and File Decryption (steps 3 and 5 in the process of the search and retrieval introduced in Section 3) onto the edge server, because they are both computation-intensive tasks. The edge server could deal with these tasks in less time because of the superiority in resources. Mobile devices are released from heavy computation tasks, which decreases the time and energy consumption in the encrypted search. What is more, they could make better use of the energy to sensor data.
In the second part, we focus on data security protection from the untrusted cloud. Our idea is to divide the encrypted data search and retrieval into two parts which have different computational components. More specifically, we assign the edge server to process the encrypted data search, and the data retrieval is processed by the cloud without changes. Therefore, the untrusted cloud cannot get any information about the keywords the user submits and the relevance between the query keyword and search results. The detailed design of how ENSURE addresses the efficiency and secure challenges will be introduced in the next subsection.
Process of File Search and Retrieval in ENSURE
To offload the encrypted data search to the edge server, the edge server should download the encrypted file index which is stored in the cloud and synchronize it with the cloud. This is entirely feasible for three reasons. Firstly, the edge server has a relatively sufficient storage space to store the file index, which is much smaller than files. Secondly, the edge server usually has a stable network condition for file index synchronization. Last but not least, the update of data is far less frequent than data search, so one synchronous interaction between the edge server and cloud may support multiple queries on average.
After the edge server downloads the file index, the workload of Index Searching (step 2 in the process of file search and retrieval introduced in Section 3) can be handled by the edge server. So, the search request is sent to the edge server but not the cloud. The data user could send the key to the edge server through a secure connection, and the key is used to decrypt the posting list related to the query keyword. The process of encrypted file search and retrieval in ENSURE is shown in Figure 3 and follows the steps:
1.
Keyword Processing: Since the data owner permits the data user to access the data, the data user could encrypt and hash the keyword to generate the search request (a trapdoor of the keyword) in their mobile device when he wants to search the top-k relevant files involving the keyword. Following this, the mobile device would send the request to the edge server and wait for the response from the edge server.
2.
Index Searching: On receiving the request, the edge server first synchronizes the encrypted file index from the cloud in case of the file index update. Then, it would decrypt the file index and search the matching posting list in the index based on the search request. The posting list contains the information (e.g., word frequency) of files that involve the keyword.
3.
Calculation&Rank: The edge server obtains and decrypts the posting list corresponding to the keyword, then uses the information in the posting list to calculate the relevance scores to find the top-k relevant files. Lastly, since all the files are still stored in the cloud storage, a request should be sent to the cloud server in order to retrieve these files.
4.
File Retrieval: The cloud finds the retrieved files and sends them back to the edge server upon request.
5.
File Decryption: The edge server receives these files from the cloud. After decryption, the top-k relevant files are sent to the mobile device.
Compared with the traditional scheme, we conclude that offloading the relevance score calculation and file decryption onto the edge server releases the user's mobile device from a heavy computation load, which is the main bottleneck to reducing the file search/retrieval time and saving energy. Moreover, the edge sever blocks the access of the curious cloud to the relevance between the query keyword and the search results because the index searching process is handled by the edge server and the curious cloud only retrieves target files upon request, but has no way to learn the keywords corresponding to these files.
Note that the edge server is a plug and play component in ENSURE, so if users cannot access the edge server, ENSURE can also support the encrypted search service with the traditional method introduced in Section 3.
Performance Efficiency of ENSURE
As our goals described, we intend to propose a solution for an encrypted search over the mobile cloud with the minimal file search/retrieval consumption in mobile devices. So, in this subsection, we theoretically discuss the improvement in time and energy consumption compared with the traditional method.
The process of file search and retrieval in two schemes is composed of five steps. These steps are listed below.
1.
Keyword Processing: In two schemes, this step makes no difference, so the execution time is equal.
2.
Index Searching: The index search time of ENSURE is faster than the traditional method because in the traditional method, the client has a round-trip communication with the cloud to send the processed keyword and retrieve the posting list corresponding to the query keyword, but ENSURE only needs the client to send the search request to the edge server for the local index and wait for the results.
3.
Calculation&Rank: The computation workload in this step is an increasing function of the document frequency (the number of files containing the keyword). In the traditional method, the high document frequency results in the rapid growth of the execution time of Calculation&Rank because the resource-constrained mobile device cannot afford this heavy workload. As for ENSURE, it allows the edge server to implement this step. With the relative abundant computing resources, the edge server is slightly effected by the high document frequency and keeps the growth of execution time relatively slow. 4.
File Retrieval: The time consumption of this step mainly depends on the network bandwidth accessing the cloud. The higher the bandwidth, the faster the speed of file retrieval. In both of the schemes, the cloud sends the relevant files back upon request. The files are only determined by the keyword. Thus, if the network condition is the same, the time consumption exhibits almost no differences between the two schemes in this step.
5.
File Decryption: In ENSURE, the file decryption is handled by the edge server, so the file decryption time of ENSURE is smaller than the traditional method which leverages the mobile device to decrypt the files. However, after decrypting the files, the edge server should send the original data back to the mobile device, which is not needed in the traditional method. The extra transmission latency is too small to disturb the efficiency improvement since the edge server is in the same LAN with mobile devices and the distance between them is quite short.
As for the performance of energy consumption, the client in ENSURE only needs to handle the Keyword Processing and afford some communication overhead such as establishing a secure connection to transmit the secret key. Thus, the superiority of ENSURE is easily observed.
Security of ENSURE
Security is also the main characteristic of our proposed design. In the subsection, we analyze the security of ENSURE compared to the traditional method in detail from the perspective of the information interaction between the cloud server and the client.
In the traditional method, two information interactions exist between the cloud server and the client, which happens in Index Searching and File Retrieval, respectively. In the process of Index Searching, the cloud server gets the keyword from the client and replies with the corresponding post list to the client. In the process of File Retrieval, the cloud server receives the request of target files and returns these files. Based on the keyword and target files, the untrusted cloud is most likely to obtain the major term of files by deducing the association between them.
In ENSURE, there is no direct information interaction between the cloud server and the client, because the edge server acts as a middleman, which blocks the cloud server from getting keywords from the client. Thus, the only information the cloud server obtains in the process is the list of top-k relevant files. Since the cloud server could not establish the connection between keywords and relevant files, the major term of files would not be leaked. Therefore, ENSURE resists the security risk from the uncontrollable and untrusted cloud.
Evaluation
In this section, after describing the experimental environment in Section 5.1, we discuss the evaluation of ENSURE performance on the file search/retrieval time (FSRT) and energy consumption in Sections 5.2 and 5.3. For comparative purposes, we also implement the traditional method in each experiment.
Experimental Environment
To evaluate the ENSURE system, we choose 5402 Request for Comments (RFC) documents with the type of TXT to be our data set. These RFC documents are collected from The Internet Engineering Task Force (IETF) [38]. The experimental testbed is comprised of a VM with a Quad vCPU; a PC with an Intel Core i5-4590 CPU (3.3 GHz, Golden Field Industrial Co., Ltd., Dongguan, China), 8 G memory, and a network with a 10 Mbps rate; and an android smartphone with a Snapdragon 810 CPU, 3 G RAM, and 10 Mbps TD-LTE network (Xiaomi Inc., Beijing, China). They play the role of the cloud server, the edge server, and the mobile device, respectively. Additionally, the smartphone installs an android application which uses the ENSURE system to search and retrieve the relevant files. As for the connection among them, the mobile device accesses the edge server via a 50 Mbps WLAN, and the edge server connects to the cloud via a high-speed 5 GHz 802.11n network. Besides the Trepn Power Profiler [39], a power profiling tool for the smartphone with a Qualcomm CPU is installed on the smartphone to observe the energy consumption. Furthermore, in terms of the implementation details of the ranked keyword search scheme, the algorithm theory in [8] is referred to as the traditional method. ENSURE adjusts the algorithm execution units based on the system architecture as mentioned in Section 4.3.
File Search and Retrieval Time
In the experiment, we want to search the top-three relevant documents for the selected keywords and measure the FSRT. To analyze the impact of the document frequency on the FSRT simultaneously, we select five keywords with different document frequencies ranging from 97 to 2137. The selected keywords and their related document frequency are shown in Table 1. The smartphone uses the application introduced previously to submit the search request of each keyword to the edge server respectively. Then, the smartphone waits for the edge server to handle the request and sends the top-three relevant files back. The process could be divided into five parts, which are process keyword, index search, calculation&rank, retrieve, and decryption, as we described in Section 4. Thus, the FSRT is equal to the sum of the time that each part takes. We measure the FSRT of each keyword in two schemes, and the experimental results are shown in Figure 4. The results show that ENSURE could saves about 15% of the FSRT for the low document frequency keyword and 49% for the high document frequency keyword. With the frequency increasing, the difference in performances among the two schemes are widened. More specially, it is the time to complete the index search, calculation&rank, and decryption steps that results in the widened difference. The three parts are all computation-intensive tasks and would increase the burdens of the smartphone when searching high document frequency words. Offloading these tasks onto the edge server which is equipped with relatively abundant computing resources benefits from reducing the FSRT. As for the time to process the keyword, the smartphone generates the trapdoor of the keyword according to the security protocol. While it is relatively easy to implement and irrelevant to the document frequency, the time to this part remains short and stable. Additionally, the time to retrieve is determined by the network bandwidth and the size of the relevant files. Since the smartphone requests the top-three relevant files and there is little difference in the size of the files, the time to retrieve is almost unchanged.
Energy Consumption
As described above, we focus on the energy consumption on the smartphone in this subsection to prove that ENSURE is more energy-saving than the traditional method. Five sets of experiments are conducted to measure the energy consumption of two schemes with different document frequency words. The results are shown in Figure 5a. From the smartphone's perspective, ENSURE enables the smartphone to reduce the energy consumption by 38-69%. The energy consumption of ENSURE fluctuates slightly with the change of document frequency, while the energy consumption of traditional methods increases sharply.
For further analyzing the energy consumption in detail, we evaluate the power versus time. The power-time diagram is illustrated in Figure 5b, and the data is collected in the case of using the keyword "Internet" to search the top-three relevant documents. As illustrated in Figure 5b, there exists four major energy consumption points in the traditional method and two in ENSURE. Through combined analysis of the FSRT in Section 5.2, we could easily find that the four points correspond to Index Searching, Calculation&Rank, File Retrieval, and File Decryption, respectively. Furthermore, they require high demands on computation resources and consume most of the energy. Since Index Searching, Calculation&Rank, and File Decryption are offloaded onto the edge server in ENSURE, the smartphone is released from heavy computation. The two major energy consumption points in ENSURE are the communication overhead with the edge server and File Retrieval; but both of them would not disturb the smartphone too much. Overall, the energy consumption of ENSURE is significantly reduced and less affected by document frequency of the keyword.
Conclusions
This paper presents ENSURE, a new architecture for supporting an efficient and secure encrypted search over mobile cloud storage. In particular, ENSURE makes use of powerful edge computing resources to handle computation-intensive tasks and deal with sensitive data. In that way, ENSURE could achieve a high-efficiency performance and minimize the information acquisition of the curious cloud. To that aim, this paper has shown how we carefully redesign the search procedure and evaluate it with the comparison of the traditional method. Our experimental results show that ENSURE could result in much faster queries and significant energy saving, especially when the document frequency of the searching keyword is high in the data set.
However, our current work still results in some possible extensions. In this paper, we take advantage of the fully trusted edge resources such as our own laptop or home PCs, but not all edge resources are fully trusted. So, it would be of significance to find a secure way to leverage these edge resources in the future. | 7,270.2 | 2018-04-01T00:00:00.000 | [
"Computer Science",
"Engineering"
] |
Cryptosporidium rubeyi n. sp. (Apicomplexa: Cryptosporidiidae) in multiple Spermophilus ground squirrel species
Previously we reported the unique Cryptosporidium sp. “c” genotype (e.g., Sbey03c, Sbey05c, Sbld05c, Sltl05c) from three species of Spermophilus ground squirrel (Spermophilus beecheyi, Spermophilus beldingi, Spermophilus lateralis) located throughout California, USA. This follow-up work characterizes the morphology and animal infectivity of this novel genotype as the final step in proposing it as a new species of Cryptosporidium. Analysis of sequences of 18S rRNA, actin, and HSP70 genes of additional Cryptosporidium isolates from recently sampled California ground squirrels (S. beecheyi) confirms the presence of the unique Sbey-c genotype in S. beecheyi. Phylogenetic and BLAST analysis indicates that the c-genotype in Spermophilus ground squirrels is distinct from Cryptosporidium species/genotypes from other host species currently available in GenBank. We propose to name this c-genotype found in Spermophilus ground squirrels as Cryptosporidium rubeyi n. sp. The mean size of C. rubeyi n. sp. oocysts is 4.67 (4.4–5.0) μm × 4.34 (4.0–5.0) μm, with a length/width index of 1.08 (n = 220). Oocysts of C. rubeyi n. sp. are not infectious to neonatal BALB/c mice and Holstein calves. GenBank accession numbers for C. rubeyi n. sp. are DQ295012, AY462233, and KM010224 for the 18S rRNA gene, KM010227 for the actin gene, and KM010229 for the HSP70 gene.
Introduction
Cryptosporidium spp. are a group of protozoan parasites that infect a wide range of vertebrate hosts including companion animals, livestock, wildlife, and humans. Approximately 30 species of Cryptosporidium have been described in vertebrate hosts that include fish, amphibians, reptiles, birds and mammals ( Slapeta, 2013). Host specificity, when documented, is highly variable between Cryptosporidium species, with some species or genotypes, for example Cryptosporidium parvum, capable of infecting multiple vertebrate hosts, while other species, for example, Cryptosporidium andersoni, appear restricted to a much smaller number of hosts. Systematic challenge studies for many recently described species of Cryptosporidium in taxonomically-related or unrelated vertebrate hosts are often lacking. Although humans and livestock are considered major biological reservoirs of a number of Cryptosporidium species (MacKenzie et al., 1994;Atwill et al., 2006;Feltus et al., 2006;Brook et al., 2009), wildlife are increasingly recognized as significant sources of environmental dissemination (Jiang et al., 2005;Feng et al., 2007;Ruecker et al., 2007;Chalmers et al., 2010) which can help foster inter-species transmission between livestock, wildlife, and humans (Hill et al., 2008;Putignani and Menichella, 2010;Raskova et al., 2013).
Ground-dwelling squirrels of the genus of Spermophilus are ubiquitous across California, USA. Each Spermophilus species inhabits a different set of ecosystems, including coastal plains and lower agricultural valleys, foothills dominated by annual grassland or oak woodlands, meadow complexes surrounded by coniferous forests, and isolated groves of pinyon pines in the remote mountains of eastern California. Colonies of ground squirrels can reach relatively high densities in suitable habitats, resulting in high rates of environmental loading of Cryptosporidium oocysts (Atwill et al., 2001). For example, California ground squirrels (Spermophilus beecheyi) can reach densities as high as 92 adults hectare À1 (Owings et al., 1977;Boellstorff and Owings, 1995), which when combined with shedding of up to 2 Â 10 5 oocysts animal À1 day À1 results in rates of environmental loading equivalent to 1 Â 10 7 oocysts hectare À1 day À1 .
Previously we have reported a unique Cryptosporidium sp. cgenotype in California ground squirrels (S. beecheyi) (Sbey03c, 05c), Belding's ground squirrels (Spermophilus beldingi) (Sbld05c), and Golden mantled ground squirrels (Spermophilus lateralis) (Sltl05c) from throughout California, USA (Pereira et al., 2010). Based on DNA sequences of multiple genes of Cryptosporidium, this c-genotype is consistently different from other Cryptosporidium isolated from a wide range of hosts, supporting its designation as a new species of Cryptosporidium in Spermophilus ground squirrels from throughout California Pereira et al., 2010). Oocysts of Sbey03c were not infectious to neonatal BALB/c mice . In the present work, we describe oocyst morphology of the c-genotype, and assess its infectivity for BALB/c mice and calves. We further characterize this genotype using 18S rRNA, actin, and HSP70 genes. The objective of the present work is to provide data on phenotypic and genotypic characteristics of c-genotype oocysts to support our assertion that this novel Cryptosporidium species in Spermophilus ground squirrels of California, USA is a new species.
Sample collection
In 2011, 100 S. beecheyi squirrels from the Central Coastal region of California were sampled for additional genetic analysis of Cryptosporidium isolates. Squirrels were collected according to the American Veterinary Medical Association's guidelines for harvesting wildlife and feces were obtained from the large intestine and colon. Fecal samples were placed into 15 ml tubes with 5 ml of antibiotic storage solution (0.1 ml 10% Tween 20, 0.006 g Penicillin G, 0.01 g Streptomycin Sulfate, 1.0 ml amphotericin B solution, and reagent grade water for a total of 100 ml). Fecal samples were placed on ice during transportation and stored at 4 C in the laboratory and processed within one week of collection.
Detection of Cryptosporidium oocysts
Detection of Cryptosporidium oocysts in previous studies were conducted by direct immunofluorescent microscopy (IFA) as described previously Pereira et al., 2010). Similar methods were used for the feces collected in 2011. Briefly, fecal samples were processed within one week after collection. Feces and antibiotic solution were mixed in deionized water with 0.2% Tween 20 to a final volume of 40 ml. The fecal suspension was strained through 4 layers of cotton gauze into a 50 ml centrifuge tube, which was filled with deionized water to a final volume of 50 ml. Tubes were centrifuged at 1500 g for 15 min and supernatant discarded, leaving a 1:1 ratio of pellet to solution volume. This final suspension was homogenized and 10 ml was used for making slides using the Aqua-Glo G/C Direct kit (Waterborne Inc., New Orleans, LA, USA). Slides were examined using a fluorescent microscope (Olympus BX 60) at Â400 magnification.
Oocyst morphology
A subset of positive fecal samples were resuspended in 40 ml of deionized water with 0.2% Tween 20 and filtered through 4-fold gauze. Filtrates were centrifuged at 1500 g for 10 min, supernatants discarded by aspiration, and the pellet resuspended with an equal volume of deionized water. Oocysts were purified using a discontinuous sucrose gradient method (Arrowood and Sterling, 1987) and washed 3Â in deionized water with centrifuging. Oocysts were counted using a phase contrast hemacytometer and concentrations were adjusted to 10 5 oocysts/ml deionized water and stored at 4 C for up to 14 days before morphology was examined. Wet mount slides were prepared by pipetting 20 ml of each oocyst stock solution on to a glass slide, applying a coverslip and sealing with nail enamel. The length and width of each oocyst were measured using Nomarski Differential Interference Contrast (DIC) microscopy (Olympus BX 60) at Â1000 magnification, with an eyepiece micrometer etched with 0.2 mm divisions (reticule KR-230, Scientific Instrument Company, Napa, CA, USA). The mean length and width and the shape index (the ratio of length to width) of each isolate were calculated based on measurements of 20 intact oocysts of each isolate. These measurements were compared to the mean shape indices of 20 oocysts of C. parvum from a naturally infected dairy calf from central California (GenBank accession no. FJ752165).
Infectivity of Cryptosporidium sp. c-genotype oocysts
An in vivo neonatal BALB/c mouse assay was used to determine if Cryptosporidium oocysts from Spermophilus squirrels were infectious for this well-studied host species. Fresh oocysts were purified as described in Section 2.3 (above) and were stored in deionized water at 4 C for approximately 3 weeks before inoculation to animals. Prior to inoculating to mice, oocysts were examined with DIC microscopy and confirmed to be intact. Female BALB/c mice with neonatal pups were purchased from Harlan Laboratories (San Diego, CA, USA), housed in cages fitted with air filters and given food and water ad libitum. Oocysts were administered to neonatal mice at 5 days of age by intragastric inoculation using a 24-gauge ballepoint feeding needle. One hour prior to infection, the pups were removed from the dam to empty their stomachs for easier inoculation and the dam was returned to the pups after inoculation. Each litter of mice was given oocysts from only one isolate as shown in Table 2, using doses ranging from 10 2 to10 4 oocysts per mouse. C. parvum oocysts (GenBank accession no. FJ752165) purified from naturally infected California dairy calves were similarly administered to mice as a positive control, as was deionized water as a negative control. Heat inactivated (incubation at 70 C for 2 h) C. parvum oocysts were also inoculated into mice to monitor passthrough of oocysts resulting from inoculation .
Cryptosporidium infection in mice was assessed by staining intestinal homogenates with a FITC-labeled anti-Cryptosporidium immunoglobulin M antibody (Waterborne Inc., New Orleans, LA, USA) which has been shown to be a sensitive method for detecting Cryptosporidium oocysts from intestinal homogenates of infected mice (Hou et al., 2004). Seven days post-inoculation (PI) mice were euthanized by CO 2 asphyxiation and the entire intestine from duodenum to rectum was collected. Intestinal samples were suspended in 5 ml of deionized water in 50 ml tubes and homogenized with an IKA ® Ultra-Turrax T8 tissue homogenizer (GmbH & Co. KG, Staufen, Germany). The homogenates were washed 1Â in deionized water by centrifuging at 1500 g for 10 min and the supernatant removed. The pellets were resuspended in 10 ml of deionized water and filtered through a 20 mm pore nylon net filter (Millipore, Bedford, MA, USA) fixed on a Swinnex holder (Millipore, Bedford, MA, USA). The filtrates were concentrated to 1 ml by centrifuging at 1500 g for 10 min and mixed by vortexing. Fifty ml of the final homogenates were mixed with 50 ml of anti-Cryptosporidium monoclonal antibodies (Waterborne Inc., New Orleans, LA, USA) and 2 ml of 0.5% evans blue, then incubated at room temperature for 45 min in a dark box. Three wet mount slides were prepared from each sample using 20 ml of reaction mixture per slide. Slides were examined with a fluorescent microscope (Olympus BX 60) and a mouse was considered infected if one or more oocysts were detected in the intestinal homogenates.
In addition to the mice infectivity assay, we also conducted a trial to measure the infectivity of the c-genotype oocysts from S. lateralis ground squirrels in two-day old Holstein calves. Newborn calves were purchased from commercial dairy farms. For each of the eight isolates of c-genotype oocysts from S. lateralis ground squirrels in Tables 1 and 2, two calves were orally inoculated, one with 100 oocysts and one with 5000 oocysts. A positive control calf was given 5000 C. parvum oocysts from dairy calves and a negative control group of two calves were not given oocysts. Fecal excretion of Cryptosporidium oocysts from calves were determined using direct immunofluorescent microscopy as described above. All animal experiments with BALB/c mice and calves were approved by the Institutional Animal Care and Use Committee (IACUC) of University of California Davis.
Multiple gene analysis of Cryptosporidium isolates from S. beecheyi
Microscopic positive fecal samples were exposed to 5 cycles of freeze (À80 C) and thaw (þ70 C) then 0.2 g was used for DNA extraction using the QIAamp DNA Stool Mini Kit (Qiagen Inc., Valencia, CA, USA) according to the manufacturer's manual. Amplification of fragments of the 18S rRNA, actin, and HSP70 genes by nested-PCR were performed using primers and cycling conditions as previously described by Xiao et al. (2000) and Jiang et al. (2005) for the 18S rRNA gene, Sulaiman et al. (2002) for the actin gene, and Sulaiman et al. (2000) for the HSP70 gene. AmpliTaq DNA polymerase (Thermo Fisher Scientific, Grand Island, NY, USA) were used for all PCR amplifications. A positive control using DNA of C. parvum isolated from calves from a dairy near Davis, CA as template and a negative control without DNA template were included in each PCR. PCR products were verified by electrophoresis in 2% agarose gel stained with ethidium bromide. Products of the secondary PCR were purified using the QIAamp DNA Mini Kit (Qiagen Inc., Valencia, CA, USA) according to the manufacturer's manual. Purified DNA was sequenced in both directions at the University of California DNA Sequencing Facility, using an ABI 3730 Capillary Electrophoresis Genetic Analyzer (Applied Biosystems Inc., Foster City, CA, USA).
Sequences were analyzed and consensus sequences were generated using the Vector NTI Advanced 11 software (Invitrogen, Carlsbad, CA, USA). Consensus sequences were compared to Cryptosporidium sequences in the GenBank using NCBI's online BLAST tool with the default algorithm parameters to target 100 sequences (http://blast.ncbi.nlm.nih.gov/) (March 12, 2015 as last day accessed). Phylogenetic analyses were conducted using Genenious Basic 5.6.5. software (Biomatters, Auckland, New Zealand). Phylogenetic relationships were inferred by using the neighbor-joining method and the Tamura-Nei genetic distance model with bootstrapping of 1000 replicates for the three genes. Depending on the availability of sequences in the GenBank, reference sequences for constructing the phylogenetic trees were selected based on: 1) sequences representing well described Cryptosporidium species (exclude synonyms) from fish, amphibians, reptiles, birds, and mammals, 2) sequences previously used by others for species description or as reference sequences, 3) sequence length (longer sequence if full sequence not available for each species; i.e. 18S rRNA gene sequences !700 bp, actin gene sequences !750, and HSP70 gene sequences !1700 bp), 4) sequences not originating from cloned PCR products due to the potential for erroneous sequence data generated from cloning PCR products (Zhou et al., 2003;Ruecker et al., 2011), and 5) previously published c-genotypes from ground squirrels (i.e. Sbey-c, Sbld-c, and Sltl-c genotypes of the 18S rRNA gene). Names and GenBank accession numbers of selected references sequences are shown in Figs. 2e4. The DNA sequences of 18S rRNA gene (GQ899206), actin gene (XM_003879845), and HSP70 gene (XM_003883591) of Neospora caninum were used as out-groups for constructing the phylogenetic tress. Sequences from S. beecheyi and all selected reference sequences were trimmed at both the 5 0 and 3 0 ends after alignment to use the same length for phylogenetic tree construction.
Statistical analysis
The mean length and mean width of oocysts from ground squirrel isolates were compared to those of C. parvum oocysts from a dairy calf using a two sample T-test and the SPSS Statistics 19 software (North Castle, NY, USA).
Oocyst morphology
Oocysts of c-genotype from the three ground squirrel species appeared spherical or ovoid, morphologically similar to C. parvum oocysts from California dairy calves. The mean (±SD) size and shape index of the isolates of oocysts from S. lateralis, S. beecheyi, and S. beldingi are shown in Table 1. The width for all oocysts from S. lateralis, S. beecheyi, and S. beldingi ground squirrels were narrower than that of C. parvum oocysts while the lengths of the majority of isolates from all three ground squirrel species were shorter than that of C. parvum oocysts (Table 1). No significant differences in oocyst size were observed among c-genotype oocysts within and between each squirrel species. Mean size were 4.67 (4.4e5.0) Â 4.33 (4.0e4.8) mm with a length/width index of 1.08 for oocysts (n ¼ 160) from S. lateralis; 4.69 (4.4e5.0) Â 4.42 (4.2e4.6) mm with a length/width index of 1.06 for oocysts (n ¼ 40) from S. beecheyi; and 4.68 (4.4e5.0) Â 4.27 (4.0e5.0) mm with a length/ width index of 1.10 for oocysts (n ¼ 20) from S. beldingi. Overall mean size of oocysts from the three ground squirrel species were 4.67 (4.4e5.0) Â 4.34 (4.0e5.0) mm with a length/width index of 1.08 (n ¼ 220). Representative differential interference contrast (DIC) photos of Sbey-c (Sbey11c) genotype oocysts from S. beecheyi collected in 2011 are shown in Fig. 1. The width and length of oocysts from the majority of Cryptosporidium species or genotypes measure between 4 and 6 microns, appear nearly spherical and have obscure internal structures (Fayer et al., 2000), thus very limited morphological characteristics are available for differentiating Cryptosporidium oocysts at the species or genotype level. Although oocyst morphology alone is not reliable for identifying Cryptosporidium species or genotypes (Fall et al., 2003), morphological analysis is an important complement to molecular and biological analysis in delineating species or genotypes of Cryptosporidium .
Oocyst infectivity of Cryptosporidium sp. c-genotype
Using C. parvum from dairy calves as a positive control, 83% (5/6) of neonatal BALB/c mice were infected after inoculation of 100 oocysts and 100% (19/19 and 17/17) after inoculation of 5000 and 10,000 oocysts, respectively (Table 2). In contrast, oocysts of all isolates of Cryptosporidium sp. c-genotype from the three ground squirrel species failed to produce detectable levels of infection in mice. Cryptosporidium oocyst infectivity in mice varies with species and genotype, inoculum size, mouse species and strain, age, and susceptibility (Finch et al., 1993;Neumann et al., 2000;Hou et al., 2004). It is well established that neonatal BALB/c mice are susceptible to C. parvum infection (Fayer, 1995;Slifko et al., 2002;Jenkins et al., 2003;Guk et al., 2004). Tarazona et al. (1998) reported that inoculation of 10 4 or more C. parvum oocysts results in 100% infection in BALB/c mice. We determined previously that the 50% infective dose (ID 50 ) for C. parvum in neonatal BALB/c mice was 70.6 oocysts (Li et al., 2005) and mice inoculated with 1000 oocysts resulted in 100% infection . Previously we have shown that inoculation up to 10 4 Sbey03c oocysts failed to infect neonatal BALB/c mice and the current results confirm this, indicating that Cryptosporidium sp. c-genotype oocysts from Spermophilus ground squirrels are not infectious to neonatal BALB/c mice-and also exhibit some degree of host specificity. In a similar study, inoculation of 10 3 Cryptosporidium oocysts from red squirrels (Sciurus vulgaris) failed to generate detectable infection in neonatal and adult CD-1 and BABL/c mice (Kv ac et al., 2008).
Intestinal homogenates coupled with fluorescent microscopy for determining C. parvum infection in neonatal mice has been shown to be significantly more sensitive than histopathology (Hou et al., 2004). In the present work, no oocysts were detected from mice inoculated with heat inactivated C. parvum oocysts, which confirmed that oocysts detected in positive control mice were not from direct inoculation and subsequent pass through but instead from patent intestinal infections. No clinical signs of cryptosporidiosis were observed in C. parvum infected mice which is not unusual given that asymptomatic cryptosporidial infections in mice have been documented previously by other investigators (Tarazona et al., 1998;Kv ac et al., 2008). Prepatent periods of Cryptosporidium infection in mice vary with species and doses of oocysts, species, age, and susceptibility of mice, with younger mice generally more Fig. 1. Cryptosporidium sp. Sbey11c oocysts from California ground squirrels (S. beecheyi). Differential interference contrast (DIC) microscopy (1000Â), bar ¼ 10 mm.
susceptible (Youssef et al., 1992;Tarazona et al., 1998;Matsui et al., 1999;Rhee et al., 1999;Yang et al., 2000). In the present work most mice were euthanized at day 7 PI for detection of oocysts, which was appropriate for the detection of C. parvum infection in mice in the present and previous work (Hou et al., 2004). To explore the possibility of a longer prepatent period for Cryptosporidium infection in mice from inoculation of Spermophilus ground squirrel oocysts, we postponed euthanasia to day 10 PI in some mice inoculated with oocysts from S. lateralis (isolates 113, 128, 155, and 230). Despite this longer period, no oocysts were detected in this cohort of mice. This suggests that the failure to detect Cryptosporidium infection in neonatal BALB/c mice was due to host specificity of Spermophilus-derived Cryptosporidium rather than the length of the prepatent period.
Although only two calves were inoculated for each of the 8 isolates, we did not find evidence of infection in calves from inoculation with up to 5000 oocysts of the c-genotype from eight S. lateralis ground squirrels; rather, calves in all groups including the negative control group (without oocyst inoculation) eventually became infected with Cryptosporidium oocysts that were confirmed to be 100% identical to the C. parvum via sequencing the 18S rRNA gene. This genotype of oocyst was the same as found in our positive control calf whereby the oocysts were collected from a local dairy in the same region where the calves were purchased (data not shown). Given that Cryptosporidium remain genetically stable after passing through mammalian species (Akiyoshi et al., 2002), these calfhood infections with C. parvum might be due to natural infection before inoculation or cross contamination from, for example, filth flies from nearby commercial dairies and/or from our positive control calves. Our calf pens were in an outdoor open facility which can allow filth flies to circulate between positive control and other calves. Our results of BALB/c mice and calf infectivity studies suggest there exists host specificity for this specific c-genotype Cryptosporidium shed by Spermophilus ground squirrels.
Multiple gene analysis of Cryptosporidium sp. c-genotype isolates from S. beecheyi
We previously reported DNA fingerprinting of Cryptosporidium isolates from Spermophilus ground squirrels collected throughout California, USA (longitude of 114 8 0 W to 124 24 0 W and latitude of 32 30 0 N to 42 N) (Pereira et al., 2010). In this present work additional fingerprinting using 18S rRNA, actin, and HSP70 genes was conducted on new Cryptosporidium isolates from S. beecheyi collected in 2011 from the Central Coastal region of California (e.g., latitude of 35 16 0 N and longitude: 120 39 0 W) to confirm our earlier findings of a new species of Cryptosporidium in this host species. Among the 100 S. beecheyi squirrel fecal samples (each from a different squirrel), 18, 14, and 3 fecal samples with oocysts were successfully sequenced for the c-genotype by using the 18S rRNA, actin, and HSP70 gene, respectively. Using our previous nomenclature based on host species, year of isolation, and genotype, in this manuscript we describe the c-genotype collected in 2011 as Sbey11c (host S. beecheyi, 2011 isolation, genotype-c). According to electrophoresis and DNA sequencing results, no positive squirrels were found to be shedding more than one genotype at a time. The GenBank accession numbers of representative c-genotype sequences are KM010224 of the 18S rRNA gene, KM010227 of the actin gene, and KM010229 of the HSP70 gene, respectively. Given the small amount of fecal sample obtained using trap and release procedures for squirrels, it can be difficult to have sufficient oocysts to successfully complete PCR and multiple gene sequencing from a single isolate from this host species, but one isolate of c-genotype was successfully sequenced for all three genes. Phylogenetic trees based on DNA sequences representing the c-genotype of the three genes were constructed and juxtaposed against reference sequences of Cryptosporidium species/genotypes selected as mentioned above (Figs. 2e4).
BLAST results (as of March 12, 2015) of DNA sequences of the three genes are shown in Table 3. With respect to the actin gene, the Sbey11c (KM010227) was not 100% identical to any Cryptosporidium sequence in the GenBank, with maximal similarity of only~93% to a Cryptosporidium sp. chipmunk genotype I (JX978270). Phylogenetic analysis of the actin gene sequences revealed similar results as the BLAST analysis in that Sbey11c did not form a distinct clade with any existing Cryptosporidium sequence in GenBank (Fig. 3). For the HSP70 gene, maximal similarity of Sbey11c (KM010229) to currently available sequences was at best only~92% similar to two isolates of Cryptosporidium sp. chipmunk genotype I (JX978275, JX978276). Similarly, the phylogenetic analysis of the HSP70 gene shows that Sbey11c did not form a distinct clade with any existing Cryptosporidium sequences (Fig. 4).
For the 18S rRNA gene, BLAST results show that the Sbey11c (KM010224) was 100% identical to Sbey05c (DQ295012) and Sbey03c (AY462233), 99.64% similar to Sltl05c (DQ295014), and 98.67% similar to Sbld05c (DQ295013) ( Table 3). Sbey05c and Sbey03c represent for the most common Cryptosporidium c-genotype from S. beecheyi squirrels collected in 2005 and 2003; Sltl05c represents the typical Cryptosporidium sp. c-genotype from S. lateralis squirrels collected in 2005; Sbld05c represents the typical Cryptosporidium sp. c-genotype from S. beldingi squirrels collected in 2005, as previously reported (Pereira et al., 2010). It is interesting that additional fingerprinting of new isolates collected in 2011consistently confirm the presence of Sbey-c genotype Cryptosporidium in S. beecheyi. Phylogenetic analysis of the 18S rRNA gene sequences revealed similar results as the BLAST analysis. The Sbey11c formed a distinct clade with Cryptosporidium isolated from all three host species (S. beecheyi, S. lateralis, S. beldingi) (Sbey05c, Sbey03c, Sltl05c, Sbld05c) compared to existing Cryptosporidium sequences (Fig. 2). In particular, BLAST and phylogenetic analyses of 18S rRNA, actin, and HSP70 genes sequences demonstrated that S. beecheyi are a mammalian host of the Sbey11c genotype, and that this unique genotype is also present in other species of the genus Spermophilus from throughout California, USA.
Focusing on the 18S rRNA gene that is commonly used for Cryptosporidium speciation (Checkley et al., 2015) and has the most sequences of described species and genotypes in the GenBank, our work over a decade (Atwill et al., 2001Pereira et al., 2010 current work) has demonstrated that Cryptosporidium sp. c-genotype is the most prevalent genotype in Spermophilus ground squirrels. Combining the evidence of the presence of novel Cryptosporidium species in Spermophilus ground squirrels from our previous Pereira et al., 2010) and current work, we propose to name the Cryptosporidium sp. c-genotype in Spermophilus ground squirrels as Cryptosporidium rubeyi n. sp. GenBank accession numbers of DNA sequences for C. rubeyi n. sp. are DQ295012, AY462233, and KM010224 for the 18S rRNA gene, KM010227 for the actin gene, and KM010229 for the HSP70 gene.
Description
Order: Eucoccidiorida. with a mean length/width index of 1.08 (n ¼ 220). Prepatent period, patent period and endogenous stages are unknown. Earlier work has documented Cryptosporidium infections in a gray squirrel (Sundberg and Ryan, 1982), fox squirrels (Current, 1989), flying squirrels (Current, 1989), and a 13-lined ground squirrel (Current, 1989). Using "Cryptosporidium" and "squirrel" as key words during a recent literature search in PubMed conducted on January 13, 2015 resulted in only a few publications. C. parvum was reported in Eurasian red squirrels (Sciurus vulgaris) in Italy (Bertolino et al., 2003); C. parvum was also reported in Siberian chipmunks (Tamias sibiricus) originated from China and found infectious to SCID mice and ICR mice (Matsui et al., 2000); Cryptosporidium muris was reported in Siberian chipmunks (Eutamias sibiricus) imported from Southeast Asia to Czech Republic and found infectious to BALB/c mice (H urkov a et al., 2003); Cryptosporidium ferret genotype and chipmunk genotype were reported in red squirrels (Sciurus vulgaris L) in Italy and no detectable infection was found in CD1 mice and BALB/c mice after inoculation 1000 oocysts (Kv ac et al., 2008). All these squirrel species belong to different genus other than Spermophilus. The only documentation of Cryptosporidium in Spermophilus genus besides ours was a report of C. parvum in spotted souslik (Spermophilus suslicus) in Poland (Kloch and Bajer, 2012). S. suslicus is a different species with distinct geographic distributions compared to Spermophilus ground squirrels in California, USA. In contrast to these sporadic detections of Cryptosporidium in different species of squirrels, we have consistently detected the Cryptosporidium sp. c-genotype in Spermophilus ground squirrels from throughout California over the past decade (Atwill et al., 2001Pereira et al., 2010 present work).
Describing a novel species of Cryptosporidium requires four attributes to be satisfied: 1) genetic characterization; 2) morphometric studies of oocysts; 3) demonstration of natural and at least some experimental host specificity; and 4) compliance with International Commission on Zoological Nomenclature (ICZN) Fayer, 2010). Combining our current work and previous works Pereira et al., 2010), we have satisfied the requirements of genetic and morphometric characteristics as well as host specificity studies similar in scope to other researchers who have established Cryptosporidium scrofarum (Kv a c et al., 2013), Cryptosporidium viatorum (Elwin et al., 2012), Cryptosporidium xiaoi (Fayer and Santín, 2009), Cryptosporidium ryanae (Fayer et al., 2008), Cryptosporidium fayeri (Ryan et al., 2008), and Cryptosporidium bovis (Fayer et al., 2005). To comply with ICZN, we provide morphological description of c-genotype oocysts (see above) and present DIC photos of Sbey11c oocysts from S. beecheyi collected in 2011 (Fig. 1).
Conclusion
Our current and previous work has demonstrated that Spermophilus ground squirrels are the hosts of a distinct Cryptosporidium sp. c-genotype. Based on the findings from these work, the cgenotype in Spermophilus ground squirrels is described as C. rubeyi n. sp. Further studies are warranted to understand the geographic distribution, environmental dissemination, and epidemiology including age and sex related prevalence of C. rubeyi n. sp. in Spermophilus ground squirrels.
Conflicts of interest
The authors declared that there is no conflict of interest. | 6,606.6 | 2015-08-24T00:00:00.000 | [
"Biology"
] |
In Vivo Response to Methotrexate Forecasts Outcome of Acute Lymphoblastic Leukemia and Has a Distinct Gene Expression Profile
Background Childhood acute lymphoblastic leukemia (ALL) is the most common cancer in children, and can now be cured in approximately 80% of patients. Nevertheless, drug resistance is the major cause of treatment failure in children with ALL. The drug methotrexate (MTX), which is widely used to treat many human cancers, is used in essentially all treatment protocols worldwide for newly diagnosed ALL. Although MTX has been extensively studied for many years, relatively little is known about mechanisms of de novo resistance in primary cancer cells, including leukemia cells. This lack of knowledge is due in part to the fact that existing in vitro methods are not sufficiently reliable to permit assessment of MTX resistance in primary ALL cells. Therefore, we measured the in vivo antileukemic effects of MTX and identified genes whose expression differed significantly in patients with a good versus poor response to MTX. Methods and Findings We utilized measures of decreased circulating leukemia cells of 293 newly diagnosed children after initial “up-front” in vivo MTX treatment (1 g/m2) to elucidate interpatient differences in the antileukemic effects of MTX. To identify genomic determinants of these effects, we performed a genome-wide assessment of gene expression in primary ALL cells from 161 of these newly diagnosed children (1–18 y). We identified 48 genes and two cDNA clones whose expression was significantly related to the reduction of circulating leukemia cells after initial in vivo treatment with MTX. This finding was validated in an independent cohort of children with ALL. Furthermore, this measure of initial MTX in vivo response and the associated gene expression pattern were predictive of long-term disease-free survival (p < 0.001, p = 0.02). Conclusions Together, these data provide new insights into the genomic basis of MTX resistance and interpatient differences in MTX response, pointing to new strategies to overcome MTX resistance in childhood ALL. Trial registrations: Total XV, Therapy for Newly Diagnosed Patients With Acute Lymphoblastic Leukemia, http://www.ClinicalTrials.gov (NCT00137111); Total XIIIBH, Phase III Randomized Study of Antimetabolite-Based Induction plus High-Dose MTX Consolidation for Newly Diagnosed Pediatric Acute Lymphocytic Leukemia at Intermediate or High Risk of Treatment Failure (NCI-T93-0101D); Total XIIIBL, Phase III Randomized Study of Antimetabolite-Based Induction plus High-Dose MTX Consolidation for Newly Diagnosed Pediatric Acute Lymphocytic Leukemia at Lower Risk of Treatment Failure (NCI-T93-0103D).
A B S T R A C T Background
Childhood acute lymphoblastic leukemia (ALL) is the most common cancer in children, and can now be cured in approximately 80% of patients.Nevertheless, drug resistance is the major cause of treatment failure in children with ALL.The drug methotrexate (MTX), which is widely used to treat many human cancers, is used in essentially all treatment protocols worldwide for newly diagnosed ALL.Although MTX has been extensively studied for many years, relatively little is known about mechanisms of de novo resistance in primary cancer cells, including leukemia cells.This lack of knowledge is due in part to the fact that existing in vitro methods are not sufficiently reliable to permit assessment of MTX resistance in primary ALL cells.Therefore, we measured the in vivo antileukemic effects of MTX and identified genes whose expression differed significantly in patients with a good versus poor response to MTX.
Methods and Findings
We utilized measures of decreased circulating leukemia cells of 293 newly diagnosed children after initial ''up-front'' in vivo MTX treatment (1 g/m 2 ) to elucidate interpatient differences in the antileukemic effects of MTX.To identify genomic determinants of these effects, we performed a genome-wide assessment of gene expression in primary ALL cells from 161 of these newly diagnosed children (1-18 y).We identified 48 genes and two cDNA clones whose expression was significantly related to the reduction of circulating leukemia cells after initial in vivo treatment with MTX.This finding was validated in an independent cohort of children with ALL.Furthermore, this measure of initial MTX in vivo response and the associated gene expression pattern were predictive of long-term disease-free survival (p , 0.001, p ¼ 0.02).
Conclusions
Together, these data provide new insights into the genomic basis of MTX resistance and interpatient differences in MTX response, pointing to new strategies to overcome MTX resistance in childhood ALL.
Introduction
Childhood acute lymphoblastic leukemia (ALL), the most common cancer in children, can now be cured in approximately 80% of patients [1,2].Pharmacogenomics aims to elucidate the genomic determinants of treatment response and why treatment fails to cure the remaining 20% of patients, many of whom have favorable prognostic features [3,4].Prior studies have provided insight into the genomic determinants of resistance to several antileukemic agents [5], but methodological constraints have precluded such genomewide studies of in vitro methotrexate (MTX) resistance.This research gap is unfortunate because MTX is widely used in the treatment of many human cancers, including essentially all treatment protocols for newly diagnosed ALL [1].
The pharmacokinetics and pharmacodynamics of MTX in ALL cells are well understood, whereas the genomic determinants of the antileukemic effects of MTX remain to be elucidated [6,7].Cellular uptake of MTX is mediated by the protein reduced folate carrier [8], whereas its efflux is mediated by ATP-binding cassette (ABC), subfamily C 1 (ABCC1) and ABCC4 [9][10][11].MTX is a tight-binding inhibitor of its primary target, enzyme dihydrofolate reductase (DHFR), which disrupts cellular folate metabolism [12].Within leukemic cells, MTX is metabolized into poly(cglutamate) forms (MTXPGs) by an adenosine triphosphate (ATP)-dependent reaction catalyzed by folylpolyglutamate synthetase [13].Compared to their monoglutamate form, MTXPGs are retained longer in cells because they are not readily effluxed by ABC transporters [14,15].MTXPGs are more potent inhibitors of other target enzymes such as thymidylate synthetase, glycinamide ribonucleotide transformylase, and aminoimidazole carboxamide transformylase.These enzymes are involved in biosynthetic pathways that are critical for DNA synthesis, DNA repair, and cell replication [14,15].Furthermore, accumulation of MTXPG has also been shown to differ among major ALL subtypes [16] and to influence treatment response and outcome in childhood ALL [17][18][19].
A more complete understanding of the mechanisms of MTX resistance in ALL cells is needed if new treatment strategies are to be developed for patients whose leukemia is resistant to this important component of ALL chemotherapy [6].Prior genomic studies of ALL chemotherapy resistance have not focused on MTX [5,20,21], because the resistance of primary ALL to MTX cannot be accurately measured by in vitro methods such as the MTT assay [22].For this reason, we used the in vivo response of newly diagnosed patients to initial single-agent MTX treatment, measured as an initial decrease in circulating ALL cells, to quantitate the antileukemic effects of MTX.We then aimed to identify genes whose expression in primary ALL cells is significantly related to the in vivo antileukemic effects of MTX.
Patients and Genetic Characterization of Leukemia Cells
A total of 293 children aged 18 y or younger with newly diagnosed ALL, enrolled on the St. Jude Total Therapy XIII and XV protocols, were included in this study (Figures 1 and S1).The investigation was approved by the Institutional Review Board at St. Jude Children's Research Hospital, and signed informed consent was obtained from parents or legal guardians before enrollment.Patient characteristics (race, sex, age, pretreatment white blood cell count [WBC PRE ], ALL subtype) were assigned by investigators at St. Jude Children's Research Hospital.Race, sex, and age were determined by questionnaire; WBC PRE , ALL subtype were determined according to the clinical protocol.The diagnosis of ALL was based on morphology, cytochemical staining, and immunophenotyping of blast cells for classification as B cell lineage or T cell lineage, as previously described [23][24][25][26][27][28].The only patients excluded were those who did not have a diagnosis of ALL, or were aged , 1 y or .18 y, or were given ALL treatment prior to referral to SJCRH.
After stratification for age, WBC PRE , immunophenotype, and sex, patients were randomized to receive initial intravenous treatments of high-dose MTX (HDMTX: 1 g/m 2 ) either as HDMTX4H (HDMTX by infusion over 4 h; n ¼ 108), as HDMTX24H (HDMTX by infusion over 24 h; n ¼ 125), or as HDMTX24HþMP (HDMTX by infusion over 24 h plus mercaptopurine [MP] 1 g/m 2 by intravenous injection; n ¼ 60).All patients who received allopurinol less than 72 h prior to HDMTX were excluded from the analyses because of potential effects on de novo purine synthesis.
Circulating leukemia cells were measured before therapy (WBC PRE ) and at day 3 following start of HDMTX treatment (WBC Day3 ), prior to the administration of other antileukemic agents.Leukocyte counts were determined with a Coulter counter (model F_STKR; Coulter, Hialeah, Florida, United States).
Isolation of ALL Blasts from Bone Marrow Aspirate
ALL blasts were obtained from bone marrow aspirates at diagnosis and 42-44 h following treatment.Samples consisted of 5-10 ml of bone marrow collected in syringes containing 800 units of heparin and kept on ice until processed.Leukemic cells were obtained by density separation over a Ficoll-Hypaque gradient and washed three times with a solution of HEPES, Hanks buffered solution, and heparin, as previously described in detail [18].
RNA Extraction and Gene Expression Profiling of Diagnostic Bone Marrow ALL Cells
Of the 293 patients treated with up-front HDMTX, 161 had sufficient diagnostic ALL cells for gene expression analysis (i.e., had sufficient leukemia cells in their diagnostic bone marrow aspirates to permit RNA isolation from 5 3 10 6 to 1 3 10 7 ALL cells).High-quality total RNA was extracted with TriReagent (MRC, Cincinnati, Ohio, United States) from cryopreserved mononuclear cell suspensions from bone marrows at diagnosis.Total RNA was processed and hybridized to the HG-U133A oligonucleotide microarray (Affymetrix, Santa Clara, California, United States).This array contains 22,215 gene probe sets, representing approximately 12,357 human genes, plus approximately 3,820 expressed sequence tag clones with unknown function [29].Following removal of probe sets with .95%absent calls, 13,488 probe sets remained.Scaled expression values of all probe sets were logarithmically transformed to stabilize variance.
Additional information on the microarray methods and results can be found at http://www.stjuderesearch.org/data/.
Intracellular MTXPGs were extracted from 42-to 44-h post-treatment bone marrow ALL cells kept in a buffered solution (Tris, EDTA, and 2-mercaptoethanol [pH 7.8]) by first boiling (100 8C for 10 min), then stored frozen at À80 8C until analysis.The HPLC separation and the radioenzymatic quantitation of MTX and six polyglutamylated metabolites (MTXPG 2-7 ) were performed as previously described [30,31].These MTXPG measurements were available for 230 patients.All results were expressed as picomoles of MTX or MTXPG per 10 9 cells.Data were available for 161 patients on both WBC DDay3 and gene expression in diagnostic bone marrow leukemia cells (Figure S1).The association between each individual probe The flow chart includes study relevant protocol information for the St. Jude Children's Research Hospital Total Therapy Protocols XIIIA, XIIIB, and XV.Specifically, from the population that received ALL treatment according to one of these three protocols, the current study included only patients who received HDMTX as initial therapy.These protocols included a randomization to determine whether patients received HDMTX or not as initial treatment, the infusion time of HDMTX, and whether MP was given after MTX (LDMTX, low-dose methotrexate).Patients with an insufficient number of ALL cells for gene expression analysis were excluded, as were patients with insufficient data on circulating ALL cells to assess response over 3 d.doi:10.1371/journal.pmed.0050083.g001set and WBC DDay3 was determined using the Pearson correlation test.Gene probe sets were rank-ordered by their p-values.Final gene probe sets were selected based on a false discovery rate (FDR) cutoff of 1.5%.
Statistical Analysis and Bioinformatics
For each patient, we computed a gene expression profile using the weighted average of the expression signals of top selected genes.The Pearson correlation coefficient between each gene's expression and WBC DDay3 was used as the weight.This weighted average of expression signals was used as the summary of the top gene expression profile for each patient.Specifically, the gene expression profile was computed according to the following formula: Gene Expression Score ¼ X 50 j¼1 weight j 3 gene expression j Where j ¼ 1 . . .50 top selected gene probe sets as listed in Table S3, and weights were determined as the correlation coefficients as listed in the same table.
We compared the top 50 gene profile with the top 100 gene profile; the correlation coefficient was 0.989 (p , 0.001, Pearson correlation test).These tests were performed using standard statistical functions in R software.
We tested 37 GenMAPP pathways and 319 Gene Ontologybiological process (GO-BP) gene groupings for association with WBC DDay3 using the ''globaltest'' method [32] implemented in the R Bioconductor package [33].This test was used to infer over-representation of specific biological pathways, and the ''geneplot'' function was applied to plot the importance of selected genes with default parameters.Multiple testing was adjusted using the Bonferroni method and the FDR according to Storey and Tibshirani [34].
Cell Cycle Distribution
The percentage of ALL cells in S-phase was determined in diagnostic bone marrow aspirates from patients for whom an adequate number of cells were available (n ¼ 154).Propidium iodide-stained DNA content was measured by flow cytometry using the Coulter EPICS V flow cytometer (Coulter Electronics, Hialeah, Florida, United States), and the computer program ModFit (Verity Software House, Verity, Topsham, Maine, United States) was used to calculate the percentages of cells in G 0 /G 1 , S, and G 2 /M phase.
Treatment Outcome Analysis
The duration of disease-free survival (DFS) was defined as the time from diagnosis until the date of leukemia relapse (event), or the last follow-up (censored).Second malignancies and death due to other reasons were censored at the time of occurrence.Treatment outcome was available in 136 patients of the 293 patients treated with HDMTX with WBC PRE and WBC Day3 measured.Of note, patients treated on protocol T15 were excluded because of short follow-up.Time was also censored at the last follow-up date if no failure was observed.Single-variable analysis using Cox proportional hazards regression, as modified by Fine and Gray [35] was used to estimate the relative risk of an event.Significant associations from the single variable analyses were further evaluated in a multiple variable analysis, which included risk classification, age, lineage, and ALL subtype in addition to WBC DDay3 and the top 50 gene expression profile.DFS curves were calculated by reversing the cumulative incidence curve, where MTX poor responders represent the top quartile, intermediate responders the middle two quartiles, and good responders the bottom quartile.
Results
Relation among WBC Day3, WBC PRE , Treatment, ALL Subtype, and MTX Metabolism Patient characteristics (race, sex, age, WBC PRE , ALL subtype) were similar among patients randomly assigned to receive HDMTX4H, HDMTX24H, or HDMTX24HþMP (p .0.13, Figure S2).Furthermore, there was no difference in WBC Day3 among the three randomized treatment groups (Table 1).The lack of differences among treatment groups coupled with our previous findings of minimal de novo purine synthesis (DNPS) inhibition and antileukemic effects of a single dose of intravenous MP [36], allowed us to analyze patients treated with HDMTX4H, HDMTX24H, and HDMTX24HþMP as a single group, to enhance the statistical power of our analyses.
WBC Day3 was significantly lower than WBC PRE (n ¼ 293, p , 0.0001) and was highly correlated with WBC PRE (p , 0.0001, r ¼ 0.79, Figure 2).Therefore, to remove the effect of the pretreatment leukemia burden (WBC PRE ) we used the WBC DDay3 corresponding to the residuals from the linear regression model of WBC Day3 versus WBC PRE (Figure 2).As indicated in Figure S3, the histogram of the residuals approximates a normal distribution, in contrast to the skewed distribution of percentage drop in WBC count from diagnosis to day 3 (WBC %drop ; p ¼ 0.1 versus p , 0.001, Kolmogorov-Smirnov test).There was a statistically significant association between WBC DDay3 and polyglutamylated MTX levels (MTXPG 2-7 ) in ALL cells (n ¼ 230, p ¼ 0.0001, r ¼ À0.25, Figure 4), with higher MTXPG 2-7 associated with greater antileukemic effect.There was not a significant relation between WBC DDay3 and ALL subtype (n ¼ 293, p ¼ 0.07).
Relation among WBC DDay3 , Gene Expression, and Pathway Analysis
Our analyses of antileukemic effects after in vivo MTX treatment and gene expression in pretreatment ALL cells identified the 50 most significant gene probe sets that were associated with antileukemic effect of MTX (WBC DDay3 , Figure 3).The FDR was less than 1.5% for these gene probe sets, and each gene had a Pearson correlation coefficient higher than 0.3 or lower than À0.3 and a p-value less than 0.001.Among these genes, the expression patterns for 21 were positively and To gain more insight into the molecular and cellular pathways related to MTX response, the global test analysis was used to determine whether the gene expression profile of different pathways retrieved from the GO-BP or GenMAPP database, were significantly associated with the antileukemic effect of MTX.As listed in Table S1, a significant association was found between WBC DDay3 and various biological pathways including those involved in cell cycle regulation, DNA repair and replication, or nucleotide metabolism.To further illustrate the influence of individual genes on the antileukemic effects of MTX within the nucleotide biosynthesis pathway, the ''gene plot'' output was used.As depicted in Figure S4, three (TYMS, DHFR, and CTPS) of the ten genes belonging to the nucleotide biosynthesis pathway were most strongly negatively associated with the MTX antileukemic effect (WBC DDay3 ).
Cellular Proliferation and MTX In Vivo Response
We were able to determine both the percentage of cells in S-phase of the cell cycle and gene expression in 154 patients (these are by ALL subtype: B-lineage hyperdiploid, n ¼ 40; Blineage other, n ¼ 47; E2A-PBX1, n ¼ 14; T-ALL, n ¼ 21; TEL-AML1, n ¼ 32; BCR-ABL, n ¼ 4; MLL-AF4, n ¼ 2).There were no significant differences in the percentage of cells in S-phase among different ALL subtypes (p ¼ 0.10, Kruskal-Wallis test).The percentage of cells in S-phase of the cell cycle was positively correlated with expression of genes involved in nucleotide biosynthesis TYMS (r ¼ 0.59, p , 0.001), DHFR (r ¼ 0.39, p , 0.001), and CTPS (r ¼ 0.21, p ¼ 0.009) (Figure S5A, Table 2), with expression of TYMS being the best marker of cell proliferation.The percentage of cells in S-phase was significantly correlated with MTX response measured as WBC DDay3 , with the higher percentage in S-phase associated with a better response (WBC DDay3 , r ¼À0.20, p ¼ 0.013; Figure S5B, Table 2).The association with percentage S-phase was similar to the top 50 gene expression profile (r ¼ À0.57p , 0.001; Figure S5C, Table 2).In contrast, the percent drop in WBC count was not significantly related to percentage of cells in S-phase (WBC %drop , r ¼ 0.034, p ¼ 0.67; Table 2).
Relation between Disease-Free Survival and Expression of CTPS, TYMS, DHFR, MTX Response (WBC %drop and WBC DDay3 ), and Top 50 Gene Expression Profile
The median follow-up of patients for this analysis was 9.1 y from diagnosis, comprising patients enrolled in St. Jude Total Therapy XIII protocol (n ¼ 136).WBC DDay3 , and TYMS and DHFR expression, were related to DFS according to Cox proportional hazards regression analyses that compared patients who remained in continuous complete remission with those who relapsed during the follow-up period.The univariable analysis of variables potentially related to DFS revealed significance for expression of TYMS (hazard ratio [HR] ¼ 0.6; p ¼ 0.008), expression of DHFR (HR ¼ 0.41; p ¼ 0.015); MTX response (WBC DDay3 , HR ¼ 21.5; p , 0.001), and the top 50 gene expression profile (HR ¼ 1.09; p ¼ 0.02; Table 3), but not for the expression of CTPS or the WBC %drop .
Furthermore, multivariable Cox regression analysis (Tables 3 and S2) that also included the conventional National Cancer Institute ALL risk criteria (i.e., ALL subtype, age, and WBC at diagnosis) revealed significance for MTX response (WBC DDay3 , HR ¼ 22.6; p ¼ 0.0046), and the expression of TYMS (HR ¼ 0.58; p ¼ 0.044) and DHFR (HR ¼ 0.31; p ¼ 0.019).The top 50 gene expression profile did not reach statistical significance in predicting relapse in the overall population when other known risk factors were included, although the trend remained evident (p ¼ 0.08, Tables 3 and S2).
Discrimination of MTX Response Using the Top 50 Gene Expression Profile and Assessment in an Independent Validation Cohort
In an independent test set of 18 additional patients who received initial HDMTX according to the St. Jude Total Therapy XV protocol, we performed gene expression analysis at diagnosis and determined WBC (ALL cell) count at diagnosis and on day 3.The gene expression profile of the top 50 genes was significantly related to the residual WBC DDay3 in this patient cohort (top 50 gene profile, p ¼ 0.0065, r ¼ 0.62, Pearson correlation; Figure 6A), thus validating the gene expression profile as predictive of the MTX response in an independent cohort of patients.
Additionally, we predicted the WBC Day3 after initial MTX treatment based on the known WBC PRE in these 18 newly enrolled patients.For that, we used either the WBC DDay3 linear regression function or the median WBC %drop developed in the original test cohort of 293 patients.The sum of the differences between the observed and the predicted WBC Day3 squared was 1.042 using the WBC DDay3 linear regression model and 3.35 using the median WBC %drop .The observed WBC Day3 values are significantly closer to the predicted values using WBC DDay3 (Figure 6B) than those based on WBC %drop (Figure 6C; p ¼ 0.0025, paired t-test), thereby further indicating that WBC DDay3 is a more accurate measure of in vivo MTX response than WBC %drop .
Discussion
The current studies have identified genes that are expressed at a significantly different level in acute lymphoblastic leukemia cells of patients who exhibit a poor in vivo response to HDMTX.High-throughput genomic approaches to assess the expression levels of RNA transcripts in cancer cells are providing new insights into pathogenesis, classification, diagnosis, stratification, and prognosis of many human cancers [23,[37][38][39].The drug resistance and gene expression profiles of leukemia cells have also been used to identify genes related to the sensitivity of ALL cells to several antileukemic agents and to forecast differences in treatment response [5,20,21].These findings have also revealed novel targets for the discovery of new agents to reverse drug resistance, such as our prior discovery of MCL1 overexpression in glucocorticoid-resistant ALL [5], and the subsequent discovery that rapamycin can down-regulate MCL1 expression and increase sensitivity of leukemia cells to dexamethasone [37,40].However, prior to the current study, there has not been a comprehensive analysis of genes related to the antileukemic effects of MTX in primary leukemia cells.
We therefore evaluated MTX response in vivo after initial therapy, because this is the only possible time to assess the antileukemic effects of MTX as a single agent in patients and because there are no reliable in vitro methods.Thus, our study focused on treatment-naive ALL, and assessed de novo resistance.This revealed that WBC DDay3 is a superior measure of in vivo MTX response when compared to the percentage drop in leukemia cells (i.e., WBC %drop ), and that WBC DDay3 was predictive of long-term DFS.Furthermore, the difference in survival cannot simply be explained by differences in MTX systemic exposure (Figure S6).
To better understand the biological basis underlying MTX response in ALL cells, we used an unbiased genome-wide approach to identify genes whose expression in primary leukemia cells in vivo was significantly related to WBC DDay3 .This process revealed 48 genes and two cDNA clones that are highly related to the in vivo MTX response (WBC DDay3 ), even after adjusting for MTXPG accumulation (n ¼ 230) (Table S3).Among those genes significantly associated with MTX response were genes involved in nucleotide metabolism (TYMS and CTPS), cell proliferation and apoptosis (BCL3, CDC20, CENPF, and FAIM3), and DNA replication or repair (POLD3, RPA3, RNASEH2A, RPM1, and H2AFX).The antileukemic effects of MTX involve inhibition of purine and pyrimidine synthesis, and the current findings indicate that interindividual differences in nucleotide synthesis influence the in vivo antileukemic effects of MTX.This finding was confirmed by a global test analysis that identified the nucleotide biosynthesis pathway as one of the most discriminating biological pathways related to MTX response.Significance of the global test was largely explained by three key genes (TYMS, DHFR, and CTPS) belonging to the nucleotide biosynthesis pathway.
Our analysis also showed that low expression of DHFR, TYMS, and CTPS was significantly correlated with poor in vivo MTX response [6,41,42].It has been shown that DHFR, TYMS, and CTPS expression is associated with critical biological processes such as DNA synthesis and cell proliferation [43,44], a finding consistent with low expression of these genes reflecting a decrease in the number of ALL cells in Sphase.As MTX selectively affects cells in the S-phase of the PLoS Medicine | www.plosmedicine.orgcell cycle [7,44,45], it is likely that low expression of these genes explains the observed association with MTX response.
To support this hypothesis, we showed that the percentage of leukemia cells in the S-phase was strongly correlated with DHFR, TYMS, and CTPS expression and with the MTX in vivo response (Figure S5, Table 2).This finding does not preclude the possibility that genetically determined high TYMS expression in ALL cells is associated with a worse prognosis in ALL, as we have previously reported [46].High TYMS expression in the current study was related to higher cell proliferation, whereas higher constitutive TYMS expression is due to a genetic polymorphism in the TYMS promoter region.After remission is achieved, higher TYMS expression due to the promoter polymorphism would connote a worse prognosis due to higher levels of the MTX target, thymidylate synthase, independent of cellular proliferation rates.
Our current data showed that low cell proliferation levels, in addition to our measure of in vivo MTX response, is an important ALL cell characteristic related to worse outcome.This result is in agreement with those of a previous study that found treatment-naı ¨ve blasts with a low proliferation rate are more resistant to several anticancer drugs in vitro [47].In the current study, the gene expression profile predicting MTX response was not associated with overall disease outcome after adjusting for other known prognostic variables in the entire study population (p ¼ 0.08), but was significantly related to DFS within high-risk patients (p ¼ 0.014).Leukemia cells of patients with high-risk ALL may intrinsically have a higher potential for poor MTX response (e.g., because of oncogenic gene fusions), in contrast to lower-risk patients whose ALL cells may acquire resistance mechanisms during the 2-3 y of therapy.Further, it is possible that patients with high-risk leukemia may be more prone to acquire resistance during therapy for various reasons (e.g., greater genetic instability in their ALL cells).
Interestingly, other known folate metabolism genes were not among the top genes, suggesting that expression of the known folate metabolism genes in pretreatment ALL cells is less important in causing de novo MTX resistance than previously thought.It may well be that these folate pathway genes are important for the acquired drug resistance that emerges during MTX treatment.It is also plausible that expression or function of these proteins is not reflected by the level of their mRNA expression in ALL cells.These possibilities merit further investigation, which is beyond the scope of the current work.
Defining the genomic determinants of ALL resistance to individual antileukemic agents is essential if the pharmacogenomics of drug resistance are to be elucidated, because the current and prior studies have shown that genes discriminating drug resistance in ALL are drug specific [5,48].To assess whether the genes we identified as related to de novo MTX resistance reflect a global resistance phenotype versus a MTXspecific effect, we compared the previously reported gene expression profiles for ALL resistance to PVAD (prednisone, vincristine, asparaginase, and daunorubicin), with the top 50 genes discriminating MTX response in the current study.This comparison revealed no overlap in the genes related to MTX resistance and the 124 genes related to prednisone, vincristine, asparaginase or daunorubicin PVAD resistance [5].This result indicates that genes identified in the current study are not a marker of general drug resistance or a global predictor of survival, rather they are specific to MTX (or perhaps other antifolates, but not all ALL chemotherapy).Furthermore, we applied our MTX gene expression profile to the publicly available German/Dutch dataset [5], and documented that the MTX gene expression profile is not related to prednisolone sensitivity in this independent patient cohort (unpublished data).
Among the 50 genes that were expressed at a significantly different level in leukemia cells of MTX good responders versus poor responders, 29 were overexpressed in the MTX poor-responders.It is plausible that these overexpressed genes would be candidate targets for small molecules or other strategies to down-regulate their function, as a means to modify MTX response.Such a strategy has already proven successful in finding agents to modify the sensitivity of ALL cells to steroids [40], and it is plausible that specific inhibitors of genes overexpressed in leukemia cells resistant to MTX could be viable targets for modulating the antileukemic effects of MTX.Likewise, strategies to invoke the expression of genes that are underexpressed in MTX poor responders could be tested for their ability to modulate MTX sensitivity.
The current study is the first, to our knowledge, to identify genes whose expression is related to in vivo MTX response in patients with newly diagnosed ALL.Our data provide new insights into the genomic basis of interpatient differences in MTX response and point to new strategies for overcoming de novo MTX resistance in childhood ALL.In addition, our data indicate that early treatment response to MTX is a significant prognostic indicator for long term DFS in children with ALL. ).We used the same cutoff for MTX response (WBC DDay3 ) as in Figure 3. Therefore, the difference in survival cannot be explained by the difference in MTX AUC .Found at doi:10.1371/journal.pmed.0050083.sg006(59 KB PDF).
Table S1. The Top Ten GenMAPP (of 37) (A) and GO-BP (of 319) (B) Pathways Associated with WBC DDay3
The column titled ''probe-set correlations with WBC DDay3 '' indicates whether most probe sets in the pathway have a positive correlation, a negative correlation or a mixture of positive and negative correlations with WBC DDay3 .Found at doi:10.1371/journal.pmed.0050083.st001(56 KB DOC).Accession Numbers MIAME-compliant primary microarray data are available through the Gene Expression Omnibus (NCBI) at http://www.ncbi.nlm.nih.gov/geo/under GSE10255 and GSM258912-GSM259072.
Editors' Summary
Background.Every year about 10,000 children develop cancer in the US.Acute lymphoblastic leukemia (ALL), a rapidly progressing blood cancer, accounts for a quarter of these childhood cancers.Normally, cells in the bone marrow (the spongy material inside bones) develop into lymphocytes (white blood cells that fight infections), red blood cells (which carry oxygen round the body), platelets (which prevent excessive bleeding), and granulocytes (another type of white blood cell).However, in ALL, genetic changes in immature lymphocytes (lymphoblasts) mean that these cells divide uncontrollably and fail to mature.Eventually, the bone marrow fills up with these abnormal cells and can no longer make healthy blood cells.As a result, children with ALL cannot fight infections.They also bruise and bleed easily and, because they do not have enough red blood cells, they often complain of tiredness and weakness.With modern chemotherapy protocols (combinations of drugs that kill the fast-dividing cancer cells but leave the normal, nondividing cells in the body largely unscathed), more than 80% of children with ALL live for at least 5 years.
Why Was This Study Done?Although this survival rate is good, some patients still die because their cancer cells are resistant to one or more chemotherapy drugs.For some drugs, the genetic characteristics of the ALL cells that make them resistant are known.Unfortunately, little is known about why some ALL cells are resistant to methotrexate, a component of most treatment protocols for newly diagnosed ALL.Methotrexate kills dividing cells by interfering with DNA synthesis and repair.Cancer cells can be resistant to methotrexate for many reasonsthey may have acquired genetic changes that stop the drug from entering them, for example.These resistance mechanisms need to be understood better before new strategies can be developed for the treatment of methotrexate-resistant ALL.In this study, the researchers have determined the response of newly diagnosed patients to methotrexate and have investigated the gene expression patterns in ALL cells that correlate with good and bad responses to methotrexate.
What Did the Researchers Do and Find?The researchers measured the reduction in circulating leukemia cells that followed the first treatment with methotrexate of nearly 300 patients with newly diagnosed ALL.They also used ''microarray'' analysis to investigate the gene expression patterns in lymphoblast samples taken from the bone marrow of 161 patients before treatment.They found that the expression of 50 genes was significantly related to the reduction in circulating leukemia cells after methotrexate treatment (a result confirmed in an independent group of patients).Of these genes, the expression of 29 was higher in patients who responded poorly to methotrexate than in patients who responded well.A ''global analysis test,'' which examined the gene expression profile of different cellular pathways in relation to the methotrexate response, found a significant association between the nucleotide biosynthesis pathway (which is needed for DNA synthesis and cellular proliferation) and the methotrexate response.Finally, patients with the best methotrexate response and the 50-gene expression profile indicative of a good response were more likely to be alive after 5 years than patients with the worst methotrexate response and the poorresponse gene expression profile.
What Do These Findings Mean?These findings provide important new insights into the genetic basis of methotrexate resistance in newly diagnosed childhood ALL and begin to explain why some patients fail to respond to this drug.They also show that the reduction in circulating leukemic cells shortly after the first methotrexate dose and a specific gene expression profile both predict the long-term survival of patients.These findings also suggest new ways to modulate sensitivity to methotrexate.Down-regulation of the expression of the genes that are expressed more highly in poor responders than in good responders might improve patient responses to methotrexate.Alternatively, it might be possible to find ways to increase the expression of the genes that are underexpressed in methotrexate poor responders and so improve the outlook for at least some of the children with ALL who fail to respond to current chemotherapy protocols.
Additional Information.Please access these Web sites via the online version of this summary at http://dx.doi.org/10.1371/journal.pmed.0050083.
The US National Cancer Institute provides a fact sheet for patients and caregivers about ALL in children and information about its treatment(in English and Spanish) The UK charity Cancerbackup provides information for patients and caregivers on ALL in children and on methotrexate The US Leukemia and Lymphoma Society also provides information for patients and caregivers about ALL The Children's Cancer and Leukaemia Group (a UK charity) provides information for children with cancer and their families MedlinePlus provides additional information about methotrexate (in English and Spanish) Trial registrations: Total XV, Therapy for Newly Diagnosed Patients With Acute Lymphoblastic Leukemia, http://www.ClinicalTrials.gov(NCT00137111); Total XIIIBH, Phase III Randomized Study of Antimetabolite-Based Induction plus High-Dose MTX Consolidation for Newly Diagnosed Pediatric Acute Lymphocytic Leukemia at Intermediate or High Risk of Treatment Failure (NCI-T93-0101D); Total XIIIBL, Phase III Randomized Study of Antimetabolite-Based Induction plus High-Dose MTX Consolidation for Newly Diagnosed Pediatric Acute Lymphocytic Leukemia at Lower Risk of Treatment Failure (NCI-T93-0103D).
MTX responses as measured by WBC Day3 , WBC PRE , and MTXPG values were logarithmically transformed to normalize their respective distributions.The Pearson correlation test was applied in order to determine the association between WBC Day3 and ALL subtype, MTXPG, and WBC PRE .The difference between WBC PRE and WBC Day3 , WBC DDay3 (the WBC residual based on the linear regression of log[WBC PRE ] change to log[WBC Day3 ] ) was determined by taking the residuals of the linear regression model of WBC Day3 versus WBC PRE , which was available for 293 patients.Specifically, MTX response is defined as: WBC DDay3 ¼ 0:492 3 logðWBC PRE Þ À logðWBC Day3 Þ þ 0:0229 We indicated ''MTX poor response'' and ''MTX good response'' in Figures 2 and 3 according to the cutoff for good responders (WBC DDay3 ,À0.14) and poor responders (WBC DDay3 .0.14), based on the bottom and top quartile of 293 patients.
Figure 1 .
Figure 1.CONSORT Flow Chart Describing Patients Enrolled in Randomized Clinical Trials at St. Jude Children's Research Hospital, from Which the Current Study Population Was DerivedThe flow chart includes study relevant protocol information for the St. Jude Children's Research Hospital Total Therapy Protocols XIIIA, XIIIB, and XV.Specifically, from the population that received ALL treatment according to one of these three protocols, the current study included only patients who received HDMTX as initial therapy.These protocols included a randomization to determine whether patients received HDMTX or not as initial treatment, the infusion time of HDMTX, and whether MP was given after MTX (LDMTX, low-dose methotrexate).Patients with an insufficient number of ALL cells for gene expression analysis were excluded, as were patients with insufficient data on circulating ALL cells to assess response over 3 d.doi:10.1371/journal.pmed.0050083.g001
Figure 4 .
Figure 4. Scatterplot of WBC DDay3 Versus Level of Total MTXPGs There is a significant correlation of WBC DDay3 with the total MTXPG level in ALL cells from 230 patients (i.e., a higher total MTXPG concentration is associated with a better in vivo MTX response) (p ¼ 0.0001, r ¼ À0.25, Pearson correlation).doi:10.1371/journal.pmed.0050083.g004
Figure S1 .
Figure S1.Flow Diagram of Data and Methods Used Gray boxes indicate data used for the analyses, white boxes intermediate data, shaded boxes data analysis method used, and n the number of patients.Found at doi:10.1371/journal.pmed.0050083.sg001(657 KB PDF).
Figure S3 .
Figure S3.Histogram of WBC Change and WBC Residuals Shown are the distributions of 293 patients for (A) WBC LogChange that is defined as log(WBC PRE ) minus log(WBC Day3 ), p ¼ 0.04, i.e., is significantly different from a normal distribution; and (B) WBC DDay3 that is the residuals of the (log[WBC Day3 ] with log[WBC PRE ]) linear regression, p ¼ 0.10, i.e., is not significantly different from a normal distribution.Found at doi:10.1371/journal.pmed.0050083.sg003(653 KB PDF).
Figure S4 .
Figure S4.Gene Plot of the Gene Ontology Nucleotide Biosynthesis Pathway The gene plot gives a bar and a reference line for each gene probe set categorized for this pathway.The bar indicates the influence of each probe set on the correlation with MTX response (WBC DDay3 ).If the height of the bar exceeds the reference line the probe set is significantly related to MTX response.Marks indicate the standard deviations by which the bar exceeds the reference line.Red indicates gene probe sets with a positive correlation and green indicates gene probe sets with a negative correlation with MTX response.Found at doi:10.1371/journal.pmed.0050083.sg004(340 KB PDF).
Figure S6 .
Figure S6.MTX Systemic Exposure Was Not Different among Responder GroupsThere was no difference (p ¼ 0.82) in MTX AUC (methotrexate area under the curve, representing MTX systemic exposure) in the groups (MTX good responder[GR] versus MTX intermediate [IR] versus MTX poor responder [PR]).We used the same cutoff for MTX response (WBC DDay3 ) as in Figure3.Therefore, the difference in survival cannot be explained by the difference in MTX AUC .Found at doi:10.1371/journal.pmed.0050083.sg006(59 KB PDF).
Figure 6 .
Figure 6.Model Performance in an Independent Test Set of Patients Relation between the in vivo MTX response (WBC DDay3 ) and top 50 gene expression profile (p ¼ 0.0065, r ¼ 0.62, Pearson correlation) for the independent validation cohort (n ¼ 18).Relation between (B) the predicted log(WBC Day3 ) using either the linear model function or (C) the median percentage drop determined in 293 patients (mean difference ¼ 0.1812946.p ¼ 0.0025, paired t-test).The regression lines in graphs (B) and (C) are based on intercept equal to zero and slope equal to one.doi:10.1371/journal.pmed.0050083.g006
Table 1 .
Patient Characteristics Were Not Different among the HDMTX Treatment Groups a Chi-square test.b Reference group.doi:10.1371/journal.pmed.0050083.t001
Table 2 .
Pearson Correlation of Selected Biological and Response Parameters with Percentage of Cells in S-Phase
Table 3 .
Univariable Hazard Analysis of the Risk of Relapse with Variables Related to Initial In Vivo MTX Response and Multivariable Cox Proportional Hazard Analyses Each Including Known Prognostic Factors (i.e., ALL Subtype, Age at Diagnosis, Risk Group) doi:10.1371/journal.pmed.0050083.t003 | 9,287.6 | 2008-04-01T00:00:00.000 | [
"Medicine",
"Biology"
] |
Advanced dwarf mongoose optimization for solving CEC 2011 and CEC 2017 benchmark problems
This paper proposes an improvement to the dwarf mongoose optimization (DMO) algorithm called the advanced dwarf mongoose optimization (ADMO) algorithm. The improvement goal is to solve the low convergence rate limitation of the DMO. This situation arises when the initial solutions are close to the optimal global solution; the subsequent value of the alpha must be small for the DMO to converge towards a better solution. The proposed improvement incorporates other social behavior of the dwarf mongoose, namely, the predation and mound protection and the reproductive and group splitting behavior to enhance the exploration and exploitation ability of the DMO. The ADMO also modifies the lifestyle of the alpha and subordinate group and the foraging and seminomadic behavior of the DMO. The proposed ADMO was used to solve the congress on evolutionary computation (CEC) 2011 and 2017 benchmark functions, consisting of 30 classical and hybrid composite problems and 22 real-world optimization problems. The performance of the ADMO, using different performance metrics and statistical analysis, is compared with the DMO and seven other existing algorithms. In most cases, the results show that solutions achieved by the ADMO are better than the solution obtained by the existing algorithms.
Introduction
Optimization occurs naturally in many human endeavors, and most human decisions go through an optimal process. Optimization is deeply rooted in many branches of science, for example, a radiation reactor system with minimal emission in physics, maximizing profit in businesses, survival of the fittest in ecology, and production line design in a manufacturing system that satisfies a set of constraints [1]. There are two established methods of solving optimization problems: the mathematical and metaheuristic approach. Each method comes with specific drawbacks; for instance, the mathematical methods are gradient-dependent, which implies that the initial starting position of the population plays a significant role in its performance [2]. The drawbacks of the two methods, coupled with the fact that global optimization problems are complex in nature, and the ease of mimicking nature's way of solving problems, have significantly contributed to the surge in the rate at which researchers are proposing nature-inspired algorithms [3]. Many aspects of nature have been a source of inspiration for developing metaheuristic algorithms. Over the years, many optimization researchers have successfully used different natural phenomena as a source of inspiration to develop metaheuristic algorithms [2]. For instance, the genetic algorithm's (GA) source of inspiration is natural selection in the theory of evolution [4]. The intelligent way birds flock together inspired the design of the particle swarm optimization (PSO) [5]. Generally, problems in various domains ranging from the traveling salesman problem [6], optimal control [7]and many more, have been solved using natureinspired metaheuristic algorithms. The research community believes the success of natureinspired metaheuristic algorithms is attributed to imitating the best ways nature solves problems.
Some authors have criticized the over-reliance on metaphor-based paradigm by the natureinspired metaheuristic algorithms [8]; however, there is consensus on the many successes recorded by these algorithms in finding solutions to complex benchmark optimization problems [9]and real-world problems in the engineering domain [10]. Like all real-world optimization problems, almost all engineering problems come with several nonlinear and complex constraints depending on the design criteria and safety rules. The optimization process of all nature-inspired metaheuristic algorithms consists of steps that mimic the problem-solving process of the natural phenomena they mimic.
No one algorithm exists that solves all optimization problems optimally, meaning each can only solve some problems optimally and others suboptimally. Hence the argument for developing a new or improved high-performance algorithm that solves specific problems. Also, many novel metaheuristic algorithm developers have cited the no-free lunch theory as a basis for regularly developing new algorithms, claiming that the proposed algorithms find better solutions for optimization problems. There is also the claim by the newly proposed algorithms of balancing exploration and exploitation to better search the problem space [11]. The claim by some metaheuristic algorithms of drawing inspiration from nature is debatable, and so is the claim of novelty and strong optimization capability.
A list of some newly proposed metaheuristic algorithms is presented in Table 1. Interested readers are referred to [3,[12][13][14][15]for a more detailed list of metaheuristic algorithms proposed within the past five decades. Also, a detailed survey of metaheuristic algorithms that outlined their components and concepts, intending to analyze their similarities and differences, is given in [16,17]. Interestingly, some of the inspirations claimed in the articles are drawn from human inventions rather than naturally occurring phenomena. For instance, the social network search (SNS) draws inspiration from the social network user's efforts to gain more popularity, a human invention rather than a naturally occurring phenomenon.
Researchers have also hybridized existing metaheuristic algorithms instead of developing an entirely new metaheuristic algorithm. So many works of literature exist that have hybridized one metaheuristic algorithm with another. Some examples include the firefly algorithm hybridized with chaos theory [18], the hybridization of ant colony strategy and harmony search scheme (HPSACO) [19], particle swarm optimizer hybridized with a variant of cuckoo search called the island-based cuckoo search, and highly disruptive polynomial mutation (iCSPM) [20], hybridization of self-assembly and particle swarm optimization (SAPSO) [21], fuzzy controllers hybridized with slime mound algorithm (SMAF) [22].
Further to the novel research outcomes resulting from the metaheuristic method and their associated hybrid or variant algorithms, the area of applicability presents more research Table 1. Some nature-inspired metaheuristic algorithms with their source of inspiration (2019-2021).
Algorithm Inspiration Reference
Group teaching optimization algorithm Group teaching mechanism [23] Black widow optimization algorithm unique mating behavior of black widow spiders. [24] Chaos Game Optimization some principles of chaos theory [25] Adolescent Identity Search Algorithm (AISA) process of identity development/search of adolescents [26] Atomic orbital search basic principles of quantum mechanics [27] A novel metaheuristic optimizer inspired by behavior of jellyfish in the ocean behavior of jellyfish in the ocean [28] Quantum dolphin swarm algorithm dolphin swarm algorithm [29] Arithmetic optimization algorithm Arithmetic operators [30] Advanced arithmetic optimization algorithm Advanced arithmetic operators [31] Ebola Optimization Search Algorithm (EOSA) Ebola virus [32,33] Golden ratio optimization method (GROM) Growth in nature using the golden ratio of Fibonacci series [34] Bald eagle search optimization algorithm bald eagle [35] Black Hole Mechanics Optimization mechanics of black holes [36] Capuchin search algorithm capuchin monkeys [37] Tiki-taka algorithm football playing style [38] Cooperation search algorithm team cooperation behaviors in modern enterprise [39] Aquila Optimizer Aquila bird [40] The Sailfish Optimizer The Sailfish group hunting [41] Social Network Search social network user's efforts to gain more popularity [42] Sine-cosine and Spotted Hyena-based Chimp Optimization Algorithm (SSC) a hybrid algorithm is developed which is based on the sine-cosine functions and attacking strategy of Spotted Hyena Optimizer (SHO) [43] Archimedes optimization algorithm law of physics Archimedes' Principle [44] Battle royale optimization algorithm a genre of digital games knowns as "battle royale." [45] Thermal Exchange Metaheuristic Optimization Algorithm Newton's law of cooling [46] African vultures optimization algorithm African vultures [47] The Red Colobuses Monkey Red Colobuses Monkey [48] Remora optimization algorithm parasitic behavior of remora [49] Red deer algorithm (RDA) Red deer [50] Pelican optimization algorithm Pelican [51] Reptile optimization algorithm Hunting crocodiles [52] Squirrel search algorithm Squirrels [53] Dwarf mongoose optimization Dwarf mongoose [54] Human Felicity Algorithm Quest for the Evolution of Human Society [55] Giraffe kicking optimization Giraffe [56] Competitive search Competition [57] Criminal search optimization algorithm Police strategies [58] Horse herding optimization algorithm Horse herd [59] prospects in the field. Optimization problems in engineering and machine learning are currently being researched, with the former having received considerable research efforts. Machine learning, specifically deep learning, has demonstrated interesting performances in image analysis [61][62][63][64]but still suffers from architectural composition resulting from combinatorial problems, which require an optimization process as a solution. Efforts to address these using heuristic methods such as in [65,66], have further revealed the complexity of the optimization problem. To remedy this, studies [32,[67][68][69][70]have approached the use of metaheuristic algorithms, or a hybrid of metaheuristic algorithms, or even some high-performing variants.
In [32], the authors applied a metaheuristic algorithm to support the selection of an optimal combination of convolutional neural network hyperparameters (CNN) to address classification problems in digital mammography and chest x-ray. Similarly, metaheuristic algorithms were employed to address the challenge of network weight optimization in [70]. Authors in [69]have also adapted metaheuristic algorithms to the evolution of neural architectures, a combinatorial problem consisting of finding the best neural network components for obtaining the best performing architecture suitable for solving a particular classification problem. In [68], the problem of feature selection for reducing classifier bottleneck was addressed using the GA metaheuristic method. The study of [67]investigated the performance of a chaotic-theory-enabled FA metaheuristic in improving the dropout regularization of deep learning models. Several other studies have investigated the use of hybrids of metaheuristic algorithms in solving object detection, segmentation, classification, and image generation and reconstruction problems. However, with new variants and highperforming hybrids of these algorithms still being researched, it further reveals that improving performance in handling optimization problems in machine learning is opening up new research frontiers. As a result, the motivation for deepening the optimization process of existing optimization algorithms through designing variants and hybrids is furthering research in metaheuristic algorithms.
Although the dwarf mongoose optimization (DMO) algorithm [54]is inspired by the foraging and social-behavioral structure of the dwarf mongoose, modeling the unique compensatory behavioral adaptations of the dwarf mongoose in DMO has led to a limitation of slow convergence due to the role the value of the alpha female plays in the updating process. Therefore, in this study, an improvement on the DMO is presented that mimics four (4) different aspects of the life of the dwarf mongoose, eliminating the limitation posed by the value of the alpha. The four social structural adaptations are modeled for the optimization process: the alpha and subordinate group, the foraging and seminomadic behavior, the predation and mound protection, and the reproductive and group splitting behavior. The study identified some major stages of activities observed in the group, namely predation, territory circuiting, reproduction, group splitting, and foraging. These processes are repeated until termination criteria are met. The proposed improved algorithm is used to solve CEC 2011 and 2017 benchmark functions, consisting of 30 classical and hybrid composite problems and 22 real-world optimization problems.
Considering the dwarf mongoose has been the source of inspiration for DMO and all the natural phenomena explaining their existence and survival, this presents a promising and improved optimization process. The research question now is: considering the competitive performance demonstrated by the DMO [54], which models only a selected phenomenon in the natural phenomenon of the dwarf mongoose, could a better and improved optimization process and performance be achieved by modeling all fundamental and existential phenomena in nature? Motivated by this research question, a detailed study of literature on dwarf mongooses was examined, and all fundamental concepts were extracted for consideration. Interestingly, critical processes and stages of the dwarf mongoose were found, which motivated the optimization process and mathematical models resulting in the proposed advanced dwarf mongoose optimization (ADMO) algorithm presented in this study. The following are the technical contributions of this study: i. A new optimization process model is designed with four stages: predation, foraging and semi-nomadism, reproduction, and group splitting.
ii. Mathematical models were developed to model each of the four stages described in (i).
iii. The optimization process design in (i) and the models in (ii) were applied to design a new variant of the DMO algorithm, namely the ADMO.
iv. Exhaustive experimentation was carried out using CEC 2017 and CEC 2011 constraint benchmark optimization functions for comparative analysis of ADMO against the base algorithm and other methods.
The rest of the paper is organized as follows: In Section 2, the dwarf mongoose optimization algorithm (DMO) is presented. Section 3 presents the advanced dwarf mongoose optimization algorithm (ADMO). The experimental setup, results, and detailed discussion are presented in Section 4. Finally, the conclusion and future work is presented in Section 5.
The dwarf mongoose optimization algorithm
This section presents an overview of the DMO, including its inspiration and the optimization processes. Also, this section is divided into two subsections to enable the smooth presentation of the various aspect of the DMO. The source of inspiration and the basic behavior of the dwarf mongoose used for the DMO are discussed in subsection one. In contrast, the implementation of the model is discussed in subsection two.
Inspiration
The DMO drew its inspiration from the dwarf mongoose, also called Helogale. They are found in areas with abundant termite mounds, rocks, and hollow trees used for hiding and protection. Africa's semidesert and savannah bush are typical habitats of dwarf mongoose. They are the smallest known African carnivore and live in a family group that is a matriarchy [71,72]. The social order of the mongoose family is such that the females and the young are ranked higher than the males and the juveniles, respectively. The division of labor and altruism within these groups is the highest that has been recorded for a mammal, and each mongoose serves as a guard, babysitter, attacking predators, or attacking conspecific intruders [73][74][75][76].
The dwarf mongoose has developed specific behavior and adaptations to survive in its natural habitat. These adaptions and behavior relate to predation avoidance and nutrition. They are not known to have a killer bite but rather a skull-crushing bite using the prey's eye for orientation. Also, no cooperative killing of large prey has been observed in the dwarf mongoose family. These adaptations restrict their prey's size and significantly affect the mongooses' social behavior and ecological adaptations to achieve individual and family nutrition [76]. The DMO is modeled after two compensatory behavioral adaptations of the mongoose, namely i. Prey size, space utilization, and group size ii. Food Provisioning.
The DMO model
The DMO [54]algorithm simulates the compensatory adaptation of the dwarf mongoose as the forage. The dwarf mongoose population is divided into the alpha group, scouts, and babysitters. Each group contributes to the compensatory behavioral adaptation, which leads to a seminomadic way of life in a territory (problem space) large enough to support the entire group. The scouting for new mounds and foraging are done simultaneously by the same group of mongooses in DMO. The optimization procedures of the proposed DMO algorithm are represented in three phases, as shown in Fig 1. The red dot signifies the alpha leading the family (blue dots) to find a food source, leaving behind the babysitters with the young (exploration). Once the food source is found, the entire group feeds extensively in the area (exploitation). The family returns intermittently to exchange babysitters and repeats the cycle.
The DMO starts by randomly initializing the candidate population and computing the fitness of each. The selection of alpha female (α) is based on Eq 1.
To update a candidate's food position, the DMO uses the expression given in Eq 2.
where phi is a uniformly distributed random number [-1,1], the peep is assumed to be the alpha female's vocalization that helps keeps the family bound together on the same path. The sleeping mound (sm) is updated after every iteration using Eq 3.
The average value of the sleeping mound sm is computed by Eq 4.
The scout group is simulated using Eq 5. The scouts must look for the new sleeping mound because the dwarf mongooses are seminomadic and never return to the previous sleeping mound. This behavior activates the exploration, and DMO models the scouting and foraging to be carried out simultaneously [76].
where, rand is a random number between [0,1], is the collective-volitive movement control parameter and M ! ¼ determines the movement of the mongoose to the new sleeping mound.
The pseudocode for the algorithm is given in algorithm listing 1 (Fig 9, S1 File).
Advanced dwarf mongoose optimization algorithm model
The section presents the advanced dwarf mongoose optimization algorithm (ADMO). The ADMO is proposed to solve the low convergence rate limitation of the DMO. This situation arises when the initial solutions are close to the optimal global solution; the subsequent value of the alpha must be small for the DMO to converge towards a better solution. The proposed improvement incorporates other social behavior of the dwarf mongoose, namely, the predation and mound protection and the reproductive and group splitting behavior to enhance the exploration and exploitation ability of the DMO. The ADMO also modifies the lifestyle of the alpha and subordinate group and the foraging and seminomadic behavior of the DMO. The optimization procedures of the proposed ADMO algorithm are represented in three phases, as shown in Fig 2. This model shows five (major) stages in the dwarf mongoose mounds. These stages are territory circuit, predation, foraging, reproduction, and group splitting. The search space of the proposed algorithm is a population of dwarf mongoose individuals initialized using Eq 6. Search for the news areas in the search space is achieved using the exploration mechanism of the algorithm. The criterion leading to the exploration phase's optimization process is conditioned on comparing foraging distance covered and territory size values. When the foraging distance exceeds the given territory size, the algorithm transits to the exploration phase; otherwise, the intensification phase is maintained. Obtaining the best solution depends on a sustained high rate of avoiding predators. Predation often weakens the quality of individuals in the search space. At the same time, avoidance of predation and increased foraging outside a territory space produces high-quality individuals in the search space.
Population initialization
The ADMO population is initialized with candidate dwarf mongooses (X), as shown in Eq (6). The population is generated stochastically between the given problem's upper bound (U) and lower bound (L).
Where n is the population size for an arbitrary dwarf mongoose mound and each x i is initializezd using Eq (7) and the position of all individuals in X in the mound is represented by (8).
x n;1 x n;2 � � � x n;dÀ 1 x n;d ð8Þ where x i,j denotes the position of the j th dimension of the i th population, n denotes the population size, and d is the dimension of the problem computed using dmp(X).
Alpha and subordinate groups
Once the population is initialized, gender-based compositional differences (M1 and M2) for male and female alpha members and alpha vector (AV �! ) for representing alpha characteristics is applied to obtain the alpha male (x alpham ) and female (x alphaf ). The best individual, say x best , in the population is used for benchmarking members of the alpha group. So that if we randomly select x i and x j from the population to represent male and female respectively and mutate them to x best , then Eqs (9) and (10) hold for the alpha male and alpha female.
where M1 and M2 represents (1+rand(0,1)) and (0.5+rand(0, 1)) (0.5+rand(0, 1)) and AV �! is computed using AV �! ¼ x best 2 . We now have n-2 individuals to partition among the subordinate and juvenile groups. The subordinate often represents the largest set of individuals in a mound, followed by the juvenile. The size of individuals in subordinate set S and juvenile set J is computed using s ¼ floor nÀ 2 3 À � and j ¼ floor nÀ 2 4 À � respectively. Their members are allocated by sorting X using their individual fitness values, and the first s are allocated to S and j allocated to J.
Foraging and semi-nomadism stage
The foraging and seminomadic nature of dwarf mongooses are motivated by the fact that food sources are scattered, requiring an extensive search by the individual to find sufficient food for itself. This foraging act often takes an intensive search over a long distance (fd), in Eq (12), which will most times be greater than territory size (ts) in Eq (13). The x alphaf is known to lead the foraging party, hence its position dmp(x alphaf ) helps to compute fd�ts. Cessation of foraging is aided by predation rate pr and birth rate br, thereby lowering energy output due to reduced energy input from nutrition and, in that case, fd<ts. This is summarized in Eq (11) which computes the new state of any individual x i in the group. Reduced space utilization leads to depleted food sources hence reduced individual fitness. In addition, the lower the group size (gs), the higher fd, while ts is computed using the summation of the age (in this case the age function expressed in Eq 13) of anal and cheek marking of all individuals in the group.
Where pr and br represent the average predation and birth rates for a mound. The position of all individuals in the group is updated after every iteration using dmp(X)+1 for each x i .
Predation and mound protection
The dwarf mongoose population suffers from terrestrial and aerial attacks, wading off the attack using a group approach. The terrestrial attack is categorized into attacks from another group of dwarf mongooses and attacks from other animals. When another group of dwarf mongoose attacks φ 1 , x alpham is credited with leading all fights, followed by the subordinates s ¼ floor nÀ 2 3 À � , juvenile j ¼ floor nÀ 2 4 À � , and x alphaf . When animals that are not dwarf mongooses attacks φ 2 the mound, only the subordinates s ¼ floor nÀ 2 3 À � and juvenile j ¼ floor nÀ 2 4 À � attack the enemy. Fatalities are often associated with the more aggressive juvenile group, thereby depleting their number in the mound group. Group fitness gf in Eq (14) and density of marking post mp in Eq (15) determines if the predator wins the group or loses to the group in the case of φ 1 attack while only gf determines their win in the φ 2 attack.
where fit represents the fitness value of the individual x i . We simulate the case of φ 1 , φ 2 , or neither of (φ 1 and φ 2 ) in every iteration, with the impact and update on the loss of a group member shown in Eq (18). We represent the loss effect using a tuple of current group members, group fitness, and the density of marking posts. We simplify Eq (18) by showing how cases 1 and 2 are computed using Eqs (16) and (17).
where a, b, k, and l denote the index of the first subordinate in the population, the index of the first juvenile in the population, the number of subordinates, and the number of juveniles, respectively, affected during an attack. Note that k must satisfy 0 � k � s where s ¼ floor nÀ 2 3 À � , and 0 � l � j where j ¼ floor nÀ 2 4 À � .
Reproduction and group splitting
The x alphaf is the only female who can raise young in a mound, rendering female subordinates and female juveniles incapable of childbearing. All cases (100%) of estruses cycle among the x alphaf leads to pregnancy, while only 62.5% of estruses cycle for subordinates leads to pregnancy. However, the young resulting from the female subordinates are either killed at birth or unable to survive since they cannot suckle. As a result, an increase in group size gs is strictly the exclusive right of x alphaf . Studies showed that the average frequency of young by the subordinate female is 0.66 compared with 9.66 for the alpha female. As a result, the reproduction (addition) of young into the population is updated using Eq (19).
where alphayoung is computed thus: alphayoung ¼ floor n�9:66 100 À � . For group splitting, dwarf mongooses are contractors rather than expansionists to preserve an economically defensible area to avoid depletion of resources (e.g., food) for the group and promote reproduction. Although group splitting is not frequent, when it does occur, the splinter group, often motivated and led by independent females, exits the mound for the main group and moves to another territory to form a new group. This often decreases gs and gf.
Because this group exit often excludes the x alpham and x alphaf The subordinate (S) members often constitute the independent female and her followers breaking away from the main group. We simulate the impact of the group splitting on group size using Eq (20.
( Individual fitness depends on the cost and benefit relationship the individual partakes in the group. Notably, the fitness value of dominant members is higher than those of the subordinates; hence two cost factors and benefit factors: � �! for the dominant group and subordinates, respectively. Meanwhile, since individual fitness sums up the group fitness, we compute the fitness and secretion (anal and cheek) of an arbitrary individual subordinate as follows in Eqs (21) and (22): The values for the vector pair CST1 �� �! ; BF1 � �! are obtained by duplicating the best and worst individuals among the dominant group and dividing both by the size of the dominant group.
Similarly, the values for the pair CST2 �� �! ; BF2 � �! are computed in the same manner except that the best and worst individuals are selected from the subordinate groups, and the division operation is done using the size of the subordinate group.
ADMO procedure
To achieve the algorithmic design of the proposed ADMO model, we first present the procedural description of the model. This is to illustrate the flow of processes in the algorithm and flowchart. The optimization strategy obtained from the dwarf mongoose begins with population initialization. This is followed by some major activities observed in the group. These activities include predation, territory circuiting, reproduction, group splitting, and foraging. These processes are repeated until termination criteria are met. A representation of the pseudocode for the algorithm is given below • Generate a defined number of dwarf mongoose individuals.
• Each dwarf mongoose belonging to each subgroup is evaluated using a domain-specific fitness function to obtain the current best individual. The current best is explicitly defined as the global best.
• Based on the fitness evaluation of all individuals, sort the population and assign individuals to the subgroups: alpha male, alpha female, subordinates, and juveniles • Initialize and set domain-specific control parameters such as Group fitness (gf), the density of marking post (mp), • For a defined number of iterations, and while the termination condition is not satisfied, REPEAT -Compute using Eqs (12) and (13) the model on the territory circuit stage -Compute using Eq (11) the model on the foraging phase to obtain foraging distance (fd) and territory size (ts) If foraging distance exceeds settlement territory size, THEN -Mongoose foraging due to depletion in food in settlement space Otherwise, -Mongoose still have food to sustain the group in the current settlement (mounds) -Derive the nature of predation by computing the values for φ 1 , φ 2 , or neither of (φ 1 and φ 2 ) If φ 1 , or φ 2 holds, THEN Check if its -Compute using the first condition on Eq (18) Otherwise -Compute using the second condition on Eq (18) otherwise -Compute using the third condition on Eq (18) -Generate a random number of young alpha species and add them to the population: reproduction/evolution phase • Using Eq (20), split the group to achieve two new dwarf mongoose groups existing independently • Compute the current best fitness and update the global best • Go up to check if the termination condition is not satisfied. Otherwise, move to the next line
• RETURN best solution
In Fig 3, a detailed procedure representation is described, with all identified model stages highlighted. In addition, we indicate where the exploration and exploitation phases of the proposed ADMO are balanced.
Computational complexity
The computational complexity of the DMO and eight (8) other algorithms are measured as defined in [77], and their results are presented in Tables 2-4. The algorithms are implemented using MATLAB R2020b, Windows 10 OS environment, Intel Core<EMAIL_ADDRESS>CPU, and 16G RAM. The time (T0) needed to run the program (D=10, 30, 50) below is measured: In the same vein, the time (T1) needed to run f18 (D=10, 30, 50) from the CEC 2017 test suit 200,000 times, and the mean time (T2) for five (5) runs of the same function is measured. The value of (T2− T1)/T0 gives the complexity of the respective algorithm. From the results in Tables 2-4, the ADMO returned the minimum values compared to the other eight (8) algorithms. Conclusively, the computational complexity of the ADMO is relatively low and easy to implement.
Conceptual advantage of the ADMO
The performance of the proposed ADMO in finding the global optimum solutions to different optimization problems can be theoretically attributed to the following: • The ADMO stochastically creates a set of candidate solutions for given optimization problems and improves these solutions using the enhanced exploratory and exploitation ability of DMO. The enhancement results from the group splitting, antipredation, and reproduction activities of the dwarf mongoose, which further mutates the candidate solutions.
• The problem search space is explored and exploited as the dwarf mongooses forage across the territory. In ADMO, the foraging depends on comparing the foraging distance and territory size, ensuring the ADMO escapes local optima.
• The ADMO also has only one parameter that can be tuned.
As listed in Algorithm 2 (Fig 10, S1 File), the following algorithm reflects the mathematical model and procedural listing for the ADMO model.
The search space of the proposed algorithm is a population of dwarf mongoose individuals initialized using Eq (6) and Lines 2-4 of Algorithm 2 (Fig 10, S1 File). Search for the news areas in the search space is achieved using the exploration mechanism of the algorithm. The criterion leading to the optimization process entering the exploration phase is conditioned on comparing the values of foraging distance covered and territory size. The algorithm transits to the exploration phase when the foraging distance exceeds the given territory size. Otherwise, the intensification phase is maintained. Obtaining the best solution depends on a sustained high rate of avoiding predators. Predation often weakens the quality of individuals in the search space. In contrast, avoiding predation and increased foraging outside a territory produces high-quality individuals in the search space.
Results and discussion
The proposed improvements of the ADMO were tested to establish performance using CEC 2011 and 2017 benchmark functions, consisting of 30 classical and hybrid composite problems and 22 real-world optimization problems. The results of ADMO for benchmark functions were compared with that of DMO and seven existing population-based metaheuristic algorithms, namely: arithmetic optimization algorithm (AOA), constriction coefficient based (PSO) and GSA (CPSOGSA), whale optimization algorithm (WOA), linear population size reduction success-history based adaptive DE (LSHADE), and covariance matrix learning with Euclidean neighborhood ensemble sinusoidal LSHADE (LSHADE-cnEpSin), LSHADE with semi-parameter adaptation hybrid with CMA-ES (LSHADESPACMA) and united multi-operator EA (UMOEA). The algorithms are carefully selected because of their track records and performance in different CEC competitions. Also, they represent different metaheuristic categories available in the literature. All the algorithms and optimization problems considered were implemented using MATLAB R2020b, and Table 5 presents the different algorithm control parameters used for the experiments. Notably, the control parameters given in Table 5 are as used in their original references. Windows 10 OS environment, Intel Core<EMAIL_ADDRESS>CPU, and 16G RAM were used to conduct the experiments. The results of 51 and 25 independent runs of each algorithm for CEC 2017 and CEC 2011, respectively, are collated using the "Best, Worst, Average, and SD" performance indicators. Further statistical analysis was carried out using mean, standard deviation, Friedman test, and Wilcoxon signed test.
CEC 2017 Benchmark test function
The results of all the algorithms used in this study are presented in this section. In addition to the performance metrics stated earlier, this study also presented the solution error measure defined as f(x)−f(x � ). The solution error gives the difference between the best result (x) found in one run of the algorithm and the globally known result f(x � ) for a specific benchmark function. Tables 6-9. Clearly, the performance of ADMO across the different dimensions is competitive. Specifically, from Table 6, the ADMO found the global optimal result for 27 benchmark functions (D=10) at least once and consistently found the optimal solution for 12 out of the 27 functions over 51 runs. It can be seen from Table 7 (D=30) that the ADMO successfully found the solution for 3 benchmark functions, 2 of which were consistent over 51 runs and 1 at least once. Looking at Table 8, the ADMO found optimal solutions for 3 benchmark functions at least once and was not consistent over the 51 runs. Generally, the ADMO showed consistent performance for the unimodal problems (f1-f3). It successfully found the solutions for D=10, 30, and 50 but none for 100 dimensions. The mean value ranges from 0 to 3.95E+02, and the standard deviation is between 0 and 7.17E+01 for 10 dimensions. For 30 dimensions, the mean and standard deviation ranges between 0 to 1.96E+03, and 0 to 2.32E+02, respectively. The mean value for 50 dimensions ranges from 9.00E-01 to 6.68E+05, and the standard deviation is between 1.87E-02 and 5.26E+06. The performance of ADMO for simple multimodal functions (f4-f10) is competitive, as seen by the number of functions it successfully found solutions for. The ADMO found solutions for 3 of the simple multimodal functions over 51 runs and 5 functions at least once for 10 dimensions. The ADMO found solutions for 1 of the simple multimodal function for 30 and 50 dimensions, respectively. For the hybrid functions (F11-F20), the ADMO successfully found solutions for all 10 functions for 10 dimensions and none for 30, 50, and 100 dimensions.
Finally, the ADMO successfully found solutions for 4 composition functions (F21-F30) in 10 dimensions, none for 30 dimensions, and 1 for 50 dimensions. In most cases, the ADMO got trapped in solutions that are very close to the global optimal solutions, as noticed in the mean value and standard deviation ranging between 0 and 6.68E+05 across all dimensions. These values are small even for the worst returned result for all dimensions considered. It can conclusively be said that the ADMO is a stable and efficient algorithm for solving the CEC 2017 benchmark problems. Also, the results across all the dimensions considered showed that the performance of ADMO slightly decreases as the dimension increases. However, it still showed stability and robustness over the different dimensions.
Comparative results for CEC 2017.
The comparative results of the ADMO and 8 other state-of-the-art algorithms on the benchmark problems with varying dimensions of 10, 30, 50, and 100 are presented in Tables 10-13. The best and standard deviation are the only two performance metrics used, and the best-returned results are marked in boldface. In addition, the 9 metaheuristic algorithms are ranked according to the scoring metric defined in CEC 2017 technical report and presented in Table 14. The Wilcoxon signed test was also performed on the results returned by the 9 algorithms across the different dimensions considered, and the results are presented in Table 15.
The LSHADE, LSHADEcnEpSin, LSHADE_SPACMA, and UMOEA came first in the different CEC competitions they entered. The performance of the proposed ADMO is compared with these algorithms and candidate representation of swarm-based (WOA, DMO) and physical-based (AOA, CPSOGSA) metaheuristic algorithms. It can be seen from the results that the proposed ADMO was very competitive with the high-performing algorithms (LSHADE, LSHADEcnEpSin, LSHADE_SPACMA, and UMOEA) across all dimensions considered. The DMO, AOA, and WOA performed poorly, failing to find optimal solutions for most benchmark problems, while the CPSOGSA performed relatively better, finding solutions for 3 functions in 10 dimensions. Generally, the performance of all the algorithms deteriorated significantly as the dimensions increased. However, the ADMO showed its stability and robustness by returning the best or most competitive solutions over all the dimensions considered. The ranking of the algorithms considered based on the scoring system defined in [77]is presented in Table 14. Clearly, the five (5) high-performing algorithms were very competitive, with score differences ranging from 0.01 to 1.23, which is very small. Overall, the ADMO ranked first, outperforming the other algorithms in 10 and 30 dimensions, respectively. The graphical representation of the scores for each algorithm is shown in Fig 4. The comparative results of all algorithms considered are tested statistically using Wilcoxon's test, which is presented in Table 15. The results are presented for each dimension (10D, 30D, 50D, and 100D). From the results, the ADMO significantly outperforms the DMO, AOA, WOA, and CPSOGSA in all four (4) dimensions considered judging by the high R+ values returned by the ADMO. Also, the ADMO, UMOEA, LSHADE_SPACMA, LSHADEcnEpSin, and LSHADE were competitive, judging by the number of ties (�) returned between their comparisons. At a significance level set at α = 0.05, the Wilcoxon's test showed a significant difference in 16 out of 28 cases, which implies that the ADMO significantly outperformed 4 out of the 9 algorithms and insignificantly outperformed the remaining 4 algorithms. In detail, the ADMO performed better, the same, less than the other algorithms considered in 138, 3, 91 out of 232 cases for 10 dimensions. In 30 dimensions, the ADMO performed better, the same, or less than the other algorithms in 171, 35, 26 out of 232 cases. Similarly, the ADMO performed better, the same, less than the other algorithms in 167, 52, 13 out of 232 cases for 50 dimensions. Finally, for 100 dimensions, the ADMO performed better, the same, less than the other algorithms in 168, 52, 12 out of 232 cases. Overall, the ADMO performed better, the same, less than the other algorithms in 644, 142, 142 out of 928 cases. Conclusively, the ADMO outperformed or was competitive in 85% of all cases. Also, Fig 5 shows the superiority of the proposed ADMO over the DMO and 7 other state-of-the-art algorithms considered across all the dimensions used in this study. The results also confirmed the searchability, stability, and efficiency of the ADMO in solving the optimization problems used in this study. The performance of ADMO was not hindered by the characteristics associated with the CEC 2017 problems, which are unimodal (separable and non-separable), multimodal (separable and non-separable), hybrid, and composite benchmark functions. This performance can be attributed to the balanced exploitation and exploration introduced by explicitly defining the predation, foraging and semi-nomadism, reproduction, and group splitting activities to carry out each optimization phase.
Furthermore, the convergence behavior of all the algorithms considered and for all dimensions is shown in Fig 6. The ADMO showed a fast convergence speed early in the iteration process for all functions. This speed slows down in the middle, especially towards the end of the iteration process. Furthermore, the convergence figure of ADMO showed that global or nearglobal solutions are attained in a smaller number of iterations for most functions. The continuous exploitation and exploration further demonstrate the scalability of the ADMO until the stop criterium is met. Table 16. It should be noted that the value of the optimal solution to these problems is not available. However, the results are discussed based on four performance metrics (best, worst, mean, and standard deviation) used to summarize the results. The results are collated over 25 independent runs for all 22 benchmark functions. The population size and other algorithm-specific metrics remained as defined in Section 4.1. it can be observed that the ADMO consistently found the same solution over the 25 independent runs of the algorithm for F4, F8, and F10; this could be the optimal solution for these functions. For the rest of the function, the solution found was not consistent over the different runs of the algorithm, but they are very close to each other, judging by the very small deviation from the mean. A conclusion can be drawn that the ADMO is an effective tool for optimizing this set of problems. Next, the ADMO is compared with other algorithms to gauge its superiority and robustness further.
Comparative results for CEC 2011.
The comparative results of ADMO with other state-of-the-art algorithms used to solve the CEC 2011 real-world problems are presented in Table 17. The results are discussed based on the mean and standard deviation returned by the respective algorithms over 25 independent runs and the same experimental conditions as detailed earlier. The LSHADE, LSHADEcnEpSin, LSHADE_SPACMA, and UMOEA came first in the different CEC competitions they entered. The performance of the proposed ADMO is compared with these algorithms and candidate representation of swarm-based (WOA, DMO), human activity (gaining-sharing knowledge (GSK) based algorithm [60]), and physical-based (AOA, CPSOGSA) metaheuristic algorithms. It can be seen from the results that the proposed ADMO was very competitive with the high-performing algorithms (LSHADE, LSHADEcnEpSin, LSHADE_SPACMA, GSK, and UMOEA) across all 22 problems considered. The DMO, AOA, and WOA performed sub-optimally, failing to find optimal solutions for most benchmark problems except F4, while the CPSOGSA performed relatively better, closely following the six high performers.
The ranking of the algorithms considered based on Friedman's test is presented in Table 18. The implication is that the smaller the mean rank, the better the performance. The null hypothesis for Friedman's test is that "there is no significant difference between the distributions of the obtained results." At a significant tolerance level set at α=0.05, the test returned a p-value=0.000 which is less than α. Therefore, reject the hypothesis. Also, the ADMO returned the least mean rank and ranked first. Closely following ADMO is LSHADEc-nEpSin, then LSHADE. The least three performing algorithms are the DMO, AOA, and WOA. The graphical representation of the performance ranking of the algorithms in CEC 2011 is shown in Fig 7. A further statistical analysis was carried out using Wilcoxon's test to show a pairwise performance comparison between ADMO and the remaining algorithms, and the results are summarized in Table 19. From the results, the ADMO significantly outperforms the UMOEA, LSHADE_SPACMA, LSHADE, DMO, AOA, WOA, and CPSOGSA in all 22 problems considered judging by the high R+ values returned by the ADMO. Also, the ADMO, LSHADEcnEp-Sin, and GSK were competitive, judging by the number of ties (�) returned between their comparisons. At a significance level set at α = 0.05, the Wilcoxon's test showed that the ADMO significantly outperformed 7 out of the 9 algorithms and insignificantly outperformed the remaining 2 algorithms. The results also confirmed the searchability, stability, and efficiency of the ADMO in solving the real-world optimization problems defined in CEC 2011 used in this study.
Furthermore, the convergence behavior of all the algorithms considered and for all 22 realworld problems is shown in Fig 8. The ADMO showed a fast convergence speed early in the iteration process for most functions except F1 and F3, which converged at the later stage of the iterations. This speed slows down in the middle, especially towards the end of the iteration process. Furthermore, the convergence figure of ADMO showed that global or near-global solutions are attained in a smaller number of iterations for most functions. The continuous exploitation and exploration further demonstrate the scalability of the ADMO until the stop criterium is met.
Summary of results
To test the effectiveness and robustness of ADMO, it is applied to solve the CEC-2017 and CEC 2011 real-parameter benchmark and real-world optimization problems, respectively. Experimental results are compared with DMO and 7 other state-of-the-art algorithms, comprising 4 algorithms that came first in different CEC competitions (LSHADE, LSHADEcnEpSin, LSHA-DE_SPACMA, and UMOEA) and three other candidate representations of other categories of metaheuristic algorithms (AOA, GSK, CPSOGSA, WOA). The performance of the algorithms a scored using the metric defined in CEC 2017 technical report and Friedman's test.
ADMO ranked first among all algorithms for CEC 2017, closely followed by UMOEA, LSHA-DE_SPACMA, and LSHADEcnEpSin. Furthermore, the obtained results were statistically analyzed using Wilcoxon's test (a non-parametric test) with a significance level of 0.05. Again, the results confirmed the superiority and competitiveness of the ADMO with the compared algorithms for all functions in the test suite. The ADMO was further used to solve the set of real-world optimization problems proposed for the CEC2011 evolutionary algorithm competition. Generally, ADMO, LSHADE, LSHADEcnEpSin, LSHADE_SPACMA, GSK, and UMOEA performed significantly better than the DMO, AOA, CPSOGSA, and WOA on most functions. The ADMO showed a fast convergence speed early in the iteration process for all functions for CEC 2017. Similarly, the ADMO also showed a fast convergence speed early in the iteration process for most functions in CEC 2011 except F1 and F3, which converged at the later stage of the iterations. This speed slows down in the middle, especially towards the end of the iteration process. Furthermore, the convergence figure of ADMO showed that global or near-global solutions are attained in a smaller number of iterations for most functions. The continuous exploitation and exploration further demonstrate the scalability of the ADMO until the stop criteria are met.
Conclusion and future work
The ADMO algorithm is an improvement of the newly developed DMO. It addresses the slow convergence due to alpha value and performs exploitation and exploration better than the original DMO. The ADMO incorporated four different social life structures of the dwarf mongoose to accomplish this. The predation and mound protection and the reproductive and group splitting behavior enhance the exploration and exploitation ability of the DMO. The ADMO also modifies the lifestyle of the alpha and subordinate group and the foraging and seminomadic behavior of the DMO. In the proposed ADMO, each candidate solution is represented by an individual dwarf mongoose in the entire population of dwarf mongooses. They cooperate as a group to carry out these different activities that have been mathematically modeled to enhance the optimization abilities of the DMO.
To test the effectiveness and robustness of the ADMO, it is applied to solve the CEC-2017 and CEC 2011 real-parameter benchmark and real-world optimization problems, respectively. Experimental results are compared with DMO and 7 other state-of-the-art algorithms, comprising 4 algorithms that came first in different CEC competitions (LSHADE, LSHADEcnEp-Sin, LSHADE_SPACMA, and UMOEA) and three other candidate representations of other categories of metaheuristic algorithms (AOA, GSK, CPSOGSA, WOA). The performance of the algorithms a scored using the metric defined in CEC 2017 technical report and Friedman's test. The ADMO ranked first among all algorithms, closely followed by the 4 high-performing algorithms (LSHADE, LSHADEcnEpSin, LSHADE_SPACMA, and UMOEA). The DMO, AOA, and WOA performed poorly across all the optimization problems considered in this study.
The ADMO is easy to implement and has been proven reliable, efficient, and robust for real parameter optimization. The ADMO, as presented, is focused on solving the single constrained continuous optimization problem. However, in future work, efforts can be made to modify the ADMO to solve constrained multi-objective optimization problems, discrete optimization problems, practical engineering optimization problems, and a host of other real- world applications. Another exciting research direction is to look at ways individual dwarf mongooses can have unique parameters and evolving intelligence capabilities. Interestingly, future research studies may focus on applying the algorithm to solve high dimensions or largescale global optimization problems. A complete parametric study of the ADMO is another useful prospective research direction. Finally, the ADMO may be hybridized with any other robust metaheuristic algorithm. | 11,018 | 2022-11-02T00:00:00.000 | [
"Computer Science",
"Mathematics"
] |
Human D-aspartate Oxidase: A Key Player in D-aspartate Metabolism
In recent years, the D-enantiomers of amino acids have been recognized as natural molecules present in all kingdoms, playing a variety of biological roles. In humans, d-serine and d-aspartate attracted attention for their presence in the central nervous system. Here, we focus on d-aspartate, which is involved in glutamatergic neurotransmission and the synthesis of various hormones. The biosynthesis of d-aspartate is still obscure, while its degradation is due to the peroxisomal flavin adenine dinucleotide (FAD)-containing enzyme d-aspartate oxidase. d-Aspartate emergence is strictly controlled: levels decrease in brain within the first days of life while increasing in endocrine glands postnatally and through adulthood. The human d-aspartate oxidase (hDASPO) belongs to the d-amino acid oxidase-like family: its tertiary structure closely resembles that of human d-amino acid oxidase (hDAAO), the enzyme that degrades neutral and basic d-amino acids. The structure-function relationships of the physiological isoform of hDASPO (named hDASPO_341) and the regulation of gene expression and distribution and properties of the longer isoform hDASPO_369 have all been recently elucidated. Beyond the substrate preference, hDASPO and hDAAO also differ in kinetic efficiency, FAD-binding affinity, pH profile, and oligomeric state. Such differences suggest that evolution diverged to create two different ways to modulate d-aspartate and d-serine levels in the human brain. Current knowledge about hDASPO is shedding light on the molecular mechanisms underlying the modulation of d-aspartate levels in human tissues and is pushing novel, targeted therapeutic strategies. Now, it has been proposed that dysfunction in NMDA receptor-mediated neurotransmission is caused by disrupted d-aspartate metabolism in the nervous system during the onset of various disorders (such as schizophrenia): the design of suitable hDASPO inhibitors aimed at increasing d-aspartate levels thus represents a novel and useful form of therapy.
INTRODUCTION
For a long time, the distribution and significance of D-amino acids were called into question as they were considered the "wrong" enantiomers of amino acids; we now know they are present in all organisms where they play different, specialized roles (Alieshkevich et al., 2018;Sasabe and Suzuki, 2018). In mammals, D-serine (D-Ser) and D-aspartate (D-Asp) attracted researchers' interest most because of their involvement in physiological processes. A schematic representation of D-Ser and D-Asp metabolism and role at the tripartite synapsis is shown in Figure 1. In the past 20 years, a number of studies focused on D-Ser since, in many areas of the mature brain, it is the preferred coagonist for synaptic N-methyl-D-aspartate (NMDA) receptors, a subtype of the glutamate ionotropic receptor family. D-Ser activates the receptor by binding to the socalled "glycine-binding site" of GluN1 subunits (alternatively to glycine) and potentiates NMDA receptormediated responses (Panatier et al., 2006;Wolosker 2007;Henneberger et al., 2010;LeBail et al., 2015;Ferreira et al., 2017), thus playing an essential role in synaptic plasticity. Accordingly, perturbation of D-Ser levels in brain, cerebrospinal fluid (CSF), and serum has been related to the pathophysiology of various neurological and psychiatric disorders (Pollegioni and Sacchi, 2010), e.g., Alzheimer's disease (Wu et al., 2004;Madeira et al., 2015;Piubelli et al., 2021), schizophrenia (Hashimoto et al., 2003;Bendikov et al., 2006), and amyotrophic lateral sclerosis (Sasabe et al., 2007;Sasabe et al., 2012). In most recent times, D-Asp was also widely investigated because of its role in the central nervous, neuroendocrine, and endocrine systems (Wolosker et al., 2000;Katane and Homma, 2011;Errico et al., 2012;Chieffi Baccari et al., 2020;Usiello et al., 2020). D-Asp stimulates postsynaptic NMDA receptors by binding to the glutamate site of GluN2 subunits, Figure 1 (Errico et al., 2012;Ota et al., 2012;Errico et al., 2015). In addition, D-Asp also stimulates metabotropic glutamate receptor 5 (mGlu5) (Molinaro et al., 2010) and presynaptic α-amino-3-hydroxy-5methylisoxazole-4-propionic acid hydrate (AMPA), mGlu5, and NMDA receptors (Cristino et al., 2015). Studies performed in animal models characterized by a persistent deregulation of D-Asp levels (i.e., DDO knockout or C57BL/6J mice chronically administered D-Asp) showed that it affects NMDA receptor-dependent processes such as synaptic transmission, plasticity, and cognition (Errico et al., 2008;Errico et al., 2011a;Errico et al., 2011b;Errico et al., 2014). D-Asp counteracts the decrease in NMDA receptor signaling observed during aging (Errico et al., 2011a,b), has beneficial effects during remyelination (DeRosa et al., 2019), and influences the process leading to Aβ40 and Aβ42 aggregation in Alzheimer's disease (D' Aniello et al., 2017). Furthermore, reduced levels of D-Asp have been reported in prefrontal cortex and striatum in patients with schizophrenia (Errico et al., 2013).
In the brain, D-Asp is mainly present in neurons (Schell et al., 1997;Errico et al., 2012): its levels are highest during the embryonic phase and in the first days of life and drop gradually in adulthood (Hashimoto et al., 1993;Errico et al., 2011a;Punzo et al., 2016). In contrast, an increase in D-Asp levels is observed in endocrine glands during postnatal and adult developmental stages, which draws a parallel with the synthesis of different hormones and of melatonin (Di Fiore et al., 2014;Chieffi Baccari et al., 2020). Notably, the progressive expression of the degradative enzyme D-aspartate oxidase (DASPO or DDO, EC 1.4.3.1) during development is responsible for the observed decrease in brain D-Asp levels during the postnatal phase: in the adult brain, D-Asp is localized inversely to DASPO (Schell et al., 1997;Errico et al., 2012).
Here, we review the main properties of hDASPO and current knowledge regarding the regulation of enzyme activity and expression, with the aim of providing a useful tool to comprehend the molecular mechanisms underlying the modulation of D-Asp levels in human tissues and to push novel therapeutic strategies targeting such a valuable target.
HUMAN D-ASPARTATE OXIDASE ISOFORMS
The UniProtKB/Swiss-Prot database (www.uniprot.org/help/ uniprotkb) predicts four different splice isoforms for the human DDO gene transcript, referred to as DDO-1-4; in contrast, no splice variants for the homologous DAO gene are reported.
The DDO-2 isoform is identical to the canonical one but lacks the central region of the transcript (encoding for residues 95-153), due to the fact that exon 4 is skipped in the original transcript sequence (Figure 3). The expression of this variant, which is 59 residues shorter (282 amino acids, 30.5 kDa), in E. coli yielded an accumulation of the recombinant protein as inclusion bodies (Setoyama and Miura, 1997); therefore, it was not characterized.
The DDO-3 isoform differs from the canonical one since it encodes a protein with 28 additional N-terminal residues (369 amino acids, 41 kDa, hDASPO_369, Figure 3), an upstream alternative start codon being recognized in the primary transcript. Notably, a BLAST search reveals that this long form is highly conserved in primates: homologous forms of mouse and beef DASPOs (albeit as partial sequences) are reported in the NCBI protein database (www.ncbi.nlm.nih. gov/protein). Worthy of note, we found very recently that the DDO-3 long isoform was differentially expressed in a cohort of individuals with Alzheimer's disease. Mass spectrometry analysis identified its N-terminal peptide in the hippocampus of female patients only, while peptides corresponding to regions common to hDASPO_341 and _369 protein isoforms were detected in samples from male and female patients and healthy subjects . Compared to the canonical shorter isoform, hDASPO_369 is characterized by a very low solubility when overexpressed in E. coli: this hampered the in-depth investigation of its biochemical properties. Nevertheless, by using a transfected U87 (human glioblastoma) cell line ectopically expressing this protein variant we observed that the additional N-terminal sequence does not affect enzyme function and cellular localization. Similar to hDAAO, both the long and the canonical hDASPO isoforms were very stable (estimated half-life ∼ 100 h). However, unlike hDAAO which is mainly degraded through the lysosome/endosome pathway (Cappelletti et al., 2014), both hDASPO variants were degraded by the ubiquitin-proteasome system upon ubiquitination .
Finally, the DDO-4 transcript isoform combines features of the DDO-2 and DDO-3 isoforms: it encodes a 310 amino acid protein (34 kDa, Figure 3) containing the additional N-terminal residues of hDASPO_369 and lacking the 59-residue-long central portion. No attempts to heterologously express this protein have been undertaken so far.
BIOCHEMICAL PROPERTIES
hDASPO_341 (the canonical isoform) was first produced in recombinant form in E. coli in 1997 with a yield of 2 mg/L and partially characterized in 2015 (Setoyama and Miura, 1997;Katane et al., 2015a). More recently, the production procedure was optimized (up to 30 mg/L) so that the biochemical properties of the enzyme could be comprehensively characterized and its structure-function relationships elucidated .
hDASPO belongs to the family of FAD-containing oxidoreductases and shows the typical properties of this group of proteins : the canonical absorbance peaks at 280, 370 and 455 nm in the oxidized form; the ability to stabilize the anionic semiquinone form of the cofactor and to covalently bind sulfite; the rapid conversion of the oxidized into the reduced form by adding the substrate under anaerobic conditions; and the high reactivity of the reduced enzyme form with molecular oxygen.
hDASPO shows good activity and stability in the 8-12 pH range (Katane et al., 2015a), differently from porcine kidney DASPO (pkDASPO, where maximal activity is constant between pH 7.5 and 9) (Yamamoto et al., 2007) and mouse DASPO (mDASPO, in which activity is constant between pH 4 and 10) (Puggioni et al., 2020). The human flavoenzyme is fully stable up to 45°C, a temperature corresponding to the optimum for its enzymatic activity. The melting temperature determined following the loss of activity is 55°C (Katane et al., 2015a), higher than the value obtained following the CD signal at 222 nm (48.8°C) , this suggesting that the alteration in secondary structure leads to the loss of enzymatic activity. Both human and mouse DASPOs are stabilized by the presence of ligands in the active site Puggioni et al., 2020). As the bovine counterpart, the holoenzyme form of recombinant hDASPO is a monomer with a molecular mass of ∼40 kDa. Differently, the apoprotein form is present in solution as an equilibrium between monomeric and trimeric species . Other mammalian DASPOs have been reported to be dimeric or tetrameric (Katane et al., 2018;Puggioni et al., 2020).
The binding of the FAD cofactor to hDASPO is characterized by a tight interaction (K d ∼ 33 nM) , similar to that observed for bovine kidney (bkDASPO) (Negri et al., 1987) and rat (rDASPO) DASPOs (Katane et al., 2018). Due to the high affinity for the flavin cofactor, hDASPO is fully present in solution as an active holoenzyme , whereas a different situation is apparent for mDASPO, which shows a 100-fold higher K d value for the cofactor, suggesting that at physiological FAD concentration (approximatively 2.5 μM in murine brain) (Decker and Byerrum, 1954) this enzyme should be present in solution in equilibrium between the active and the inactive forms (Katane et al., 2018). The interaction with the cofactor modulates hDASPO conformation (altering both the secondary and the tertiary structure) and stabilizes the protein (increasing the melting temperature of ∼7°C) .
Concerning the substrate specificity, different results are reported in the literature, probably due to the different experimental conditions and assays used. All mammalian DASPOs are highly specific for acidic D-amino acids. hDASPO shows the highest activity on D-Asp and NMDA, followed by D-Glu and D-Asn (13 and 10% of the value assayed on D-Asp) and shows negligible activity on D-His and D-Pro (1% vs. D-Asp) . pkDASPO and bkDASPO prefer NMDA to other acidic D-AAs ( Table 1). All mammalian DASPOs show similar apparent K m values for D-Asp, the only exception being mDASPO, which shows a higher figure. The apparent k cat and catalytic efficiency values for hDASPO are significantly higher than those of all mammalian DASPOs (Table 1), i.e., hDASPO shows a 5-fold higher specific activity on D-Asp than the bovine enzyme.
The oxidative deamination of acidic D-amino acids catalyzed by hDASPO follows a ternary-complex mechanism in which the complex between the reduced flavin and the imino acid reacts with oxygen before the imino acid is released . During the reductive half-reaction, substrate dehydrogenation proceeds by direct transfer of a hydride from the α-carbon of the substrate to the flavin N5, as elegantly demonstrated for DAAO by Sandro Ghisla's lab (Pollegioni et al., 1997;Umahu et al., 2000;Harris et al., 2001;Saam et al., 2010). Conversion of the D-amino acid into the planar imino acid together with the flavin reduction is very fast and seems to be reversible, with an equilibrium constant of ∼5 for the overall process . The rate constant for flavin reduction by D-Asp (k red value estimated ∼1,550 s −1 ) is higher than the k cat value (230 s −1 ), suggesting that substrate oxidation is not the rate-limiting step during catalysis. However, the high apparent K d for D-Asp (23 mM) indicates that the substrate binding largely controls the reaction rate. Differently from bkDASPO (in which the rate-limiting step is represented by a conformational change related to the binding of a second molecule of D-Asp to the reduced enzyme) (Negri et al., 1988), the rate-determining step in hDASPO is represented by the reoxidation of the reduced flavin. This step corresponds to a single exponential process with a rate constant of ∼1 × 10 5 M −1 s −1 .
STRUCTURAL PROPERTIES
The 3D structure of hDASPO was solved at 3.2 Å resolution from the diffraction data collected for the C141Y/C143G variant; these TABLE 1 | Comparison of apparent kinetic parameters of mammalian DASPOs on selected substrates (determined at 21% oxygen saturation). In parenthesis is reported the specific activity value (as U/mg protein). substitutions are present in the pkDASPO sequence and do not alter the functional and structural properties of the enzyme. The structure was solved by molecular replacement using the coordinates of hDAAO as starting model (pdb 3g3e); electron density at the active site was modeled as a glycerol molecule, a component of the protein buffer . hDASPO belongs to the DAAO-like family of the α,β-protein class according to the SCOP classification system (SCOP family 4,000,124, class 1,000,002) and to the Pfam family of the FADdependent oxidoreductases (Pfam: PF01266) (Andreeva et al., 2014;Mistry et al., 2021). According to DALI analysis, the overall tertiary structure of hDASPO is similar to that of FAD-dependent oxidases active on amino acids or amines: the closest proteins are hDAAO (RMSD 1.4 Å), glycine oxidase (RMSD 2.8 Å), and sarcosine oxidase (chain B, RMSD 2.5 Å) (Holm, 2020), Table 2. The tertiary structure of hDASPO can be divided into two large domains, each formed by noncontiguous sequence regions: the FAD-binding domain (FBD) and the substrate-binding domain (SBD), Figure 4A. The FBD possesses the canonical dinucleotide binding fold (i.e., the Rossman fold containing the corresponding Wierenga consensus sequence at the N-terminus) and an alternative of the Peroxisomal Targeting Signal one sequence (Ser-Asn-Leu, PROSITE, PS00342) at the C-terminus (Wierenga et al., 1986;Setoyama and Miura, 1997). The SBD is characterized by a large mixed β-sheet formed by eight β-strands, a fold similar to that observed in other amino acid oxidases (Fraaije and Mattevi, 2000;Umhau et al., 2000;Mortl et al., 2004).
The FAD cofactor binds to hDASPO in an extended conformation, with the isoalloxazine ring located at the interface between the FBD and the SBD ; this conformation is typical of flavoproteins belonging to the glutathione reductase two family (Dym and Eisenberg, 2001). The cofactor is kept in place by several electrostatic or polar interactions (mainly H-bonds) formed with atoms of 35 residues of the protein (located within 4.5 Å from the FAD molecule). In particular, the isoalloxazine ring interacts with the protein by five electrostatic interactions (with the side chain of Ser312 and the backbone of Ala48, Met50, and Ser308) and two van der Waals contacts (with Val203 and Gly276) . The positive N-terminal dipole of the α-helix A11, belonging to the FBD, points toward the O2 position of the pyrimidine ring of the isoalloxazine and stabilizes the negative charge of the reduced cofactor delocalized in this region during catalysis. hDASPO shares the overall mode of substrate binding with the other flavin oxidases active on amino acids (e.g., DAAO, LAAO and L-amino acid deaminase) Molla et al., 2017;Ball et al., 2018). The active site is located in a deep cavity and is characterized by three positively charged arginine side chains (Arg216, Arg237, and Arg278) surrounding the substrate D-Asp, thus forming a tight network of electrostatic and H-bond interactions ( Figures 4B,C). Arg278 forms a two-point interaction with the α-COOH group of the substrate: this essential residue is conserved in all known oxidases active on amino or hydroxy acids Umena et al., 2006;Sacchi et al., 2012). Arg216 forms a salt bridge with the c-COOH group of the side chain of D-Asp: it is fundamental in shaping the substrate scope of hDASPO. As a consequence, hDASPO shows a very narrow substrate specificity in comparison with other amino acid oxidases (Job et al., 2002;Rosini et al., 2017;Molla et al., 2020). The role of Arg216 in substrate selectivity is also supported by the biochemical properties of the R216Q hDASPO variant, which shows a lower kinetic efficiency on D-Asp and gains the ability to oxidize D-Ala . The third arginine (Arg237) is located at the entrance of the active site, close to His54. These two residues, whose side chains show multiple conformations in the hDASPO molecules of the crystal asymmetric unit, play three fundamental roles: i) they act as an active site gate switching between an open and a close conformation, thus affecting the kinetics of substrate binding and product release during turnover ( Figure 5); ii) in the open conformation, they form a surface cluster of positive charges (together with Arg216) that attracts the negatively charged D-Asp substrate Molla et al., 2020); and iii) in the close conformation, they contribute to bind the substrate D-Asp: their replacement with alanine residues resulting in a decrease in kinetic efficiency of the enzyme (in particular, a ∼20-fold increase in the apparent K m for D-Asp was observed for the R237A variant) .
The residues Tyr223 and Tyr225 complete the tight network of interactions between the substrate D-Asp and the residues of the active site ( Figures 4B,C).
INHIBITORS
In principle, it should be possible to modulate levels of D-Asp in the central nervous system by blocking hDASPO using selective inhibitors. Here, activating NMDA receptor function by increasing D-Asp levels may provide a novel therapeutic approach for psychiatric diseases such as schizophrenia.
We recently demonstrated that olanzapine, a commonly used second-generation antipsychotic drug, inhibits hDASPO activity: an IC 50 value of 23.4 μM was determined, a figure not affected by FAD concentration (at 4 vs. 20 μM) . Additional first-generation (chlorpromazine and haloperidol) and second-generation (clozapine) antipsychotics and antidepressants (amitriptyline, bupropion and fluoxetine) did not affect hDASPO activity. The chronic administration of olanzapine (5 mg/kg) increased extracellular D-Asp levels in the prefrontal cortex of freely moving mice while the chronic administration of 5 mg/kg clozapine did not. A significant increase in extracellular D-Asp levels was apparent in control mice chronically treated with 5 mg/kg olanzapine while no change was apparent in DDO −/− mice .
TISSUE AND CELLULAR LOCALIZATION
DASPO is ubiquitously expressed in mammals: the highest amounts of the enzyme have been detected in kidney, liver, and CNS (Hamilton 1985;van Veldhoven et al., 1991); interestingly, the same distribution was reported for the orthologous DAAO (Pollegioni et al., 2007 and references therein). The presence and the physiological role of these two flavoenzymes in the brain prompted the investigation of their localization in different brain areas, tissues and cell populations.
DASPO distribution in the human CNS was first investigated by the group of Dariush Fahimi (Zaar et al., 2002). hDASPO was shown to occur in several brain regions and, differently from hDAAO which was mainly reported in glial cells (Verrall et al., 2007;Sasabe et al., 2014), to be dominantly expressed in neurons. hDASPO was widely distributed in cerebral cortex, hippocampus, diencephalon, brainstem, cerebellum, spinal cord, choroid plexus, and striatum (although here only in a small population of magnocellular neurons). Conversely, based on studies in rodents, the mammalian DAAO has been traditionally considered as a hindbrain enzyme, highly expressed in the cerebellum, spinal cord, and brainstem (Horiike et al., 1994;Kapoor and Kapoor, 1997). Only recently was hDAAO discovered to be more widespread than expected when its presence and activity (albeit at quite low levels) were detected in forebrain regions (Verrall et al., 2007(Verrall et al., and, 2010Madeira et al., 2008;Sasabe et al., 2014). Benzoate treatment of mild cognitive impairment patients was recently studied by resting-state functional magnetic resonance imaging scans and regional homogeneity (ReHo) analysis: the change in working memory positively correlated with decreased ReHo in right precentral gyrus and right middle occipital gyrus (Lane et al., 2021). This paper explored regional relationships between hDAAO inhibition and NMDA receptor activity enhancement in brain.
Differently from DAAO, the mammalian DASPO was detected in the endocrine system as well (pineal and pituitary glands, adrenal gland, thyroid gland, and testis), where it has been suggested to be involved in D-Asp homeostasis during adulthood and in hormone maturation . The widespread expression of the DDO gene encoding hDASPO was confirmed by consulting the GTEx Portal (www.gtexportal.org): the highest transcript levels are reported in adrenal gland, heart, brain (in particular, basal ganglia and spinal cord, while very low levels are indicated in cerebellum, hypothalamus, and brain cortex), and liver, while low levels are apparent in kidney. Low but significant levels of hDASPO transcript are also reported in female reproductive organs (but not in male ones), lung, small intestine, and spleen (Figure 6). At the transcript level, in contrast, hDAAO is essentially expressed in the cerebellum, spinal cord, liver, and kidney and is generally more abundant than hDASPO.
At the protein level, the Human Proteins Atlas (www. proteinatlas.org) further highlights the extremely different pattern of expression and tissue distribution between the two amino acid oxidases.
REGULATION OF EXPRESSION
The mammalian DASPO displays a peculiar temporal distribution pattern in the brain: studies in rat showed that protein expression and enzymatic activity are nearly absent during embryonic development, but sharply increase after birth, reaching significant expression levels during adulthood (van Veldhoven et al., 1991). In particular, postnatal expression of the enzyme is regulated at the transcriptional level by epigenetic events: the dynamic changes in DDO gene transcription are closely associated with a progressive demethylation of its putative promoter region, at eight CpG residues/islands surrounding the transcription start site (Punzo et al., 2016;Errico et al., 2020). Other work conducted at Usiello's laboratory further explored the methylation state of the DDO gene promoter among different cell types and brain areas, at FIGURE 6 | Levels of the transcripts encoding hDASPO and hDAAO in human tissues. Violin plot representing DDO (ENSG00000203797.9) and DAO (ENSG00000110887.7) gene transcript levels in different human tissues. Data were obtained from the Genotype-Tissue Expression (GTex) Project Portal, GTEx Analysis Release V8 (dbGaP Accession phs000424.v8.p2, www.gtexportal.org).Values are shown in TPM (transcripts per kilobase million); tissues in which DDO transcript values were lower than 2 TPM were omitted. Box plots represent the median and 25th and 75th percentiles. Dots are data points above or below 1.5-fold the interquartile range, displayed as outliers.
Frontiers in Molecular Biosciences | www.frontiersin.org June 2021 | Volume 8 | Article 689719 9 various developmental stage in mouse and, by analyzing specific combinations of methylated CpG sites (defined as epialleles), highlighted that neurons, oligodendrocytes, astrocytes, and microglial cells display a cell type-specific methylation pattern . Ultradeep methylation analysis of the mouse DAO gene promoter region identified dynamic demethylation at two CpG sites only in the cerebellum and in a specific time window (P1-P15) (Cuomo et al., 2019), and a specific and reproducible rearrangement of the epiallele frequency distribution in undifferentiated embryonic mouse stem cells upon neural differentiation .
The status of DDO promoter methylation and corresponding transcription levels were also investigated in the human brain (dorsolateral prefrontal cortex, hippocampus, and cerebellum of healthy subjects and schizophrenia-affected individuals) (Keller et al., 2018), showing a tissue-specific distribution of methylated CpGs with no alterations in the detected profile between diagnosis groups. The same study reported a high methylation of the DAO gene promoter region in all the investigated brain areas, consistent with the very low expression of hDAAO in the cortex and hippocampus, but clashing with the high expression levels detected in the cerebellum. Furthermore, the different brain areas are characterized by distinctive methylation signatures in both healthy and schizophrenic subjects, possibly reflecting differences in epiallele distribution among cells and a specific CpG combinatorial code (Keller et al., 2018).
The post-transcriptional regulation of DASPO expression has been investigated less and little is known about the encoding transcript stability and protein synthesis levels. Intriguingly, a preliminary in silico analysis of the 3′-UTR sequence of both DDO and DAO genes very recently disclosed the occurrence of several potential binding sites for microRNAs that are expressed in the CNS and that are known to bind sequences conserved in mammals . However, only few of the predicted binding sites were concomitantly present in human and rodents, weakening their potential as gene expression regulatory mechanisms. Moreover, the lack of post-translational modification studies on hDASPO represents a current research gap.
SINGLE NUCLEOTIDE POLYMORPHISMS AND MISSENSE VARIANTS
In their study, the work of Usiello and his group proposed that two intronic SNPs within the human DDO gene, namely, rs2057149 (A/G) and rs3757351 (C/T), have a functional role in modulating hDASPO levels in the prefrontal cortex of healthy individuals (Errico et al., 2014). The two SNPs are associated with DDO mRNA expression: the A and C alleles predict reduced transcription compared to the G and T alleles, respectively. Moreover, based on brain imaging analysis, the authors proposed that this genetically mediated decrease in hDASPO expression, mapped on the prefrontal phenotype, suggests that neuronal plasticity increased and neural networks were activated during working memory processing. Accordingly, the C allele of rs3757351 is also associated with a greater prefrontal gray matter volume and higher activity during this task than is the T allele (Errico et al., 2014). Additionally, the human gene database GeneCards (https://www.genecards.org/) reports three SNPs in the DDO gene coding sequence that present missense substitutions with no clinical significance: rs17621, rs17622, and rs17623 coding for the His230Tyr, Gln189Glu, and Leu255Arg substitutions, respectively (numbering refers to the canonical protein isoform, hDASPO_341). Interestingly, Gln189 and Leu255 are highly conserved in mammalian DASPOs but not in those of rodents, while His230 is a poorly conserved residue. These residues are located on the protein surface and the indicated substitutions probably only poorly affect the enzyme properties.
Conversely, the missense variant VAR_036244 (Phe136Leu in hDASPO_341) is reported in GeneCards as a somatic mutation in a breast cancer sample. The substituted residue is part of the β-strand 131-141 and is strictly conserved in all mammalian DASPOs. Finally, the catalogue of somatic mutations in cancer (COSMIC, https://cancer.sanger.ac.uk/cosmic) reports another missense variant as being associated with esophageal cancer, i.e., the Gly337Arg substitution in the hDASPO_369 protein isoform (corresponding to Gly309 in hDASPO_341 isoform). This Gly residue belongs to a short amino acidic sequence (HHYGHGSGG) strictly conserved in all mammalian DASPOs (and hDASPO protein isoforms). Notably, both Phe136 and Gly309 are close to the active site ( Figure 5A): in particular, the Gly309Arg substitution might deeply impact enzyme function. The COSMIC database reveals several hDAAO somatic variants identified in tumor tissues: the actual effects of the reported amino acidic substitution have not been investigated yet.
A query using the ClinVar tool at the NCBI site (www.ncbi. nlm.nih.gov/clinvar/) indicates that there are no clear relationships among missense mutations in DDO gene and human phenotypes associated with other human pathologies. All the additional SNPs currently present at the UCSC Genome Browser site (https://genome.ucsc.edu/) introduce sequence variations at noncoding regions of the DDO gene: they represent 3′-UTR, intron, and 5′-UTR variants whose effect is unknown.
D-ASPARTATE OXIDASE VS. D-AMINO ACID OXIDASE
From a biochemical point of view, hDASPO and hDAAO significantly differ in several aspects, suggesting that, during evolution, two different ways emerged to modulate D-Asp and D-Ser levels in the brain. hDASPO and hDAAO show a different pH profile while activity and stability dependence on temperature are similar (Katane et al., 2015a;Murtas et al., 2017). In solution, hDASPO is monomeric while hDAAO is a homodimer. A dimeric quaternary arrangement (similar to the one observed in hDAAO) is apparent only when hDASPO is packed into the crystal asymmetric unit: under these conditions, the stabilization of the oligomeric form is due to the presence of a phosphate and a Tris ion that mediate the monomer-monomer interactions at the dimer interface located in the apical region of the SBD. The space occupied by the phosphate and the Tris ions at the interface between monomers in hDASPO is filled by the side chains of two residues (Phe133 and Lys211) directly involved in proteinprotein contacts in hDAAO . The different oligomeric state between hDASPO and hDAAO in solution is due to differences at the dimerization interface: Asp73, His80, Phe90, Arg120, Met124, Phe133, and Lys211 residues in hDAAO are replaced by Asn73, Asn80, His90, Ala122, Lys126, Ala135, and Glu212 in hDASPO. This results in an alteration of the overall hydrophobicity and polarity of this region. Notably, the different oligomerization state could affect subcellular trafficking and the modulation by interacting proteins (Kawazoe et al., 2006;Molla et al., 2006;Molla et al., 2020). According to the manifold physiological roles, hDAAO shows a wide substrate acceptance: the best substrates are hydrophobic and bulky D-amino acids, but it is also active on small, uncharged ones. Conversely, hDASPO is highly specific for acidic substrates. At both active sites, the α-carboxylate of the substrate interacts with a Tyr and an Arg residue (Tyr223 and Arg278 in hDASPO; Tyr228 and Arg283 in hDAAO) while the side chain is accommodated in a pocket with a different chemical environment (Kawazoe et al., 2006;Molla et al., 2020). In hDASPO, this active-site pocket is positively charged and constituted by loop 217-221 and His54, Arg216, and Arg237; the same pocket in hDAAO is made up of bulky and hydrophobic residues (Leu51, Gln53, Leu215, and Ile230) (Kawazoe et al., 2006). Moreover, the conformational change of loop 216-228 and, in particular of Tyr224, allows hDAAO to bind larger substrates (Kawazoe et al., 2006;Molla et al., 2006). On the other hand, this loop is six residues shorter (loop 217-221) in hDASPO and only partially shapes the active site entrance ( Figure 5). Indeed, two additional loops in this region (namely, 54-63 and 101-107) show a significant conformational difference in comparison with hDAAO. These differences contribute to increasing the hDASPO turnover number compared to hDAAO . Actually, both flavoenzymes follow a ternary complex mechanism but with a different rate-limiting step in catalysis: it is the product release in hDAAO (<1 s −1 , since the conformational change of the active site lid is relatively slow) (Molla et al., 2006) and the step of reoxidation of the reduced flavin in hDASPO, see above . In line with this evidence, hDASPO shows on D-Asp a 10-fold higher specific activity than hDAAO on D-Ser (Molla et al., 2006;Molla et al., 2020). K m and K d values for the best substrates are in the millimolar range for both flavoenzymes, suggesting that under physiological conditions the enzymatic activity is strongly affected by the substrate concentrations (Molla et al., 2006;Molla et al., 2020). The variations in the active site architecture are also responsible for the lower affinity observed for hDASPO with CBIO and DPPD in comparison to hDAAO (Table 3): a partial steric hindrance from Ser308 and Ile52 should hamper the positioning of the rigid aromatic ring of the inhibitor parallel to the isoalloxazine ring of FAD .
No substantial differences in the mode of interaction of the cofactor with the protein moiety are apparent between the two human flavoproteins (Molla et al., 2006;Molla et al., 2020). The sole significant difference is the replacement of Trp185 in hDAAO by Gly186 in hDASPO in the region of the protein close to the FAD ribose moiety. Thus, the different affinity for the cofactor between the two flavoenzymes is probably due to different dynamics during the flavin binding and release processes. hDASPO does not seem to be modulated in vivo by ligand and flavin binding: its interaction with FAD is very strong (K d in the nanomolar range vs. a figure of 8.0 μM for hDAAO) . Owing to the weak interaction with FAD, hDAAO is present in solution as an equilibrium of holo-and apoprotein forms at the concentration of the cofactor in brain (2-5 μM) (Caldinelli et al., 2010;Murtas et al., 2017), as also reported for mDASPO (Puggioni et al., 2020), while hDASPO is fully present as active holoenzyme .
CONCLUSION AND FUTURE PERSPECTIVES
The D-enantiomer of aspartate presents many of the signatures of a classical neurotransmitter and in recent years has attracted attention for its involvement in main physiological processes. Here, recent investigations of the structure-function relationships in hDASPO are shedding light on its properties and highlighting the differences with the homologous flavoenzyme hDAAO, too. Notwithstanding a high degree of sequence and structure conservation, evolution diverged to produce amino acid oxidases that control the brain concentration of the neuromodulators D-Asp and D-Ser differently. While D-Ser levels must be maintained in a physiological range to avoid NMDA receptor hypofunction (thus, the weak activity of hDAAO is only sufficient to avoid an accumulation of D-Ser), the D-Asp levels seem to be tightly controlled by hDASPO, as observed during neurodevelopment. Such a high enzymatic efficiency might be responsible for negative outcomes since the dysfunctional D-Asp metabolism occurring during neurodevelopment may affect early critical processes related to NMDA receptors (Errico et al., 2018). For example, disrupted D-Asp metabolism (increased levels in the prefrontal cortex, hippocampus, and serum) was recently reported in BTBR mice, an animal model of idiopathic autism spectrum disorder (Nuzzo et al., 2020). A dysfunction in NMDA receptor-mediated neurotransmission due to decreased D-Asp levels in the nervous system is thought to occur during the onset of various mental disorders, including schizophrenia Errico et al., 2018). In this regard, treatment aimed at increasing D-Asp levels (and thus at activating NMDA receptor function) represents a novel and useful therapy. Such a relevant goal can be reached by acting on hDASPO: the inhibition of the enzyme activity would prevent D-Asp degradation. Known hDASPO inhibitors possess a low in vitro potency and thus may represent lead compounds for the development of new drugs based on a rational, structure-guided design, taking into account the differences in the active site geometry with hDAAO. Such molecules can be also useful to treat infertility since D-Asp is thought to be involved in the quality control of germ cells and to stimulate myelin repair in the disability associated with multiple sclerosis.
A main controversy in the field is related to the biosynthetic D-Asp pathway in human brain: the presence of a synthetic enzyme is strongly considered (both an enzyme acting similarly to serine racemase for D-Ser or using a completely different mechanism) but its identity is still elusive, which will attract the attention of researchers in the near future. Indeed, forthcoming studies also need to focus on the modulation of hDASPO function by post-translational modifications, protein interaction, and cell targeting under both physiological and pathological conditions.
AUTHOR CONTRIBUTIONS
LP conceived the manuscript. All authors wrote and critically reviewed the manuscript.
FUNDING
The work of LP, SS and GMo was supported from Fondo di Ateneo per la Ricerca. | 8,171.8 | 2021-06-23T00:00:00.000 | [
"Biology"
] |
Potential and challenges of a solid-shell element for the macroscopic forming simulation of engineering textiles
. Finite element (FE) forming simulation offers the possibility of a detailed analysis of the deformation behaviour of engineering textiles during forming processes, to predict possible manufacturing effects such as wrinkling or local changes in fibre volume content. The majority of macroscopic simulations are based on conventional two-dimensional shell elements with large aspect ratios to model the membrane and bending behaviour of thin fabrics efficiently. However, a three-dimensional element approach is necessary to account for stresses and strains in thickness direction accurately, which is required for processes with a significant influence of the fabric’s compaction behaviour, e.g. wet compression moulding. Conventional linear 3D-solid elements that would be commercially available for this purpose are rarely suitable for high aspect ratio forming simulations. They are often subjected to several locking phenomena under bending deformation, which leads to a strong dependence of the element formulation on the forming behaviour [1]. Therefore, in the present work a 3D hexahedral solid-shell element, based on the initial work of Schwarze and Reese [2,3], which has shown promising results for the forming of thin isotropic materials [1], is extended for highly anisotropic materials. The advantages of a locking-free element formulation are shown through a comparison to commercially available solid and shell elements in forming simulations of a generic geometry. Additionally, first ideas for an approach of a membrane-bending-decoupling based on a Taylor approximation of the strain are discussed, which is necessary for an accurate description of the deformation behaviour of thin fabrics.
Introduction
Lightweighting is a development strategy that aims to increase a system's efficiency and to decrease CO2 emissions rather than just reducing the weight of a system. Continuously fibre-reinforced composites made of engineering textiles like woven or noncrimp fabrics, provide a particularly high lightweight potential and have the advantage to be tailor-made for the specific application [4]. However, a detailed understanding of the manufacturing process is necessary to fully exploit this potential. Therefore, finite element (FE) forming simulations can be used to analyse the deformation behaviour and to predict effects like fibre orientation or wrinkling.
State of the art forming simulation approaches mainly apply conventional two-dimensional shell approaches [5] to efficiently model the deformation of thin fibrous reinforcements, since they allow for high aspect ratios while still accurately describing the membrane and bending behaviour. However, those approaches cannot model out-of-plane compaction, which is an important forming mechanism to predict local fibre volume contents [6] and influences the permeability of the reinforcement during processes such as wet compressions moulding [7,8]. To predict these effects and the final thickness of the part, a three-dimensional FEformulation is necessary.
The problem with utilizing most commercially available solid elements for composite forming simulations is the occurrence of numerical locking phenomena, which cause a too high bending stiffness and significantly influence the forming behaviour especially for high aspect ratios [1]. Therefore, so-called solid-shell elements have been developed, which combine techniques like reduced-integration and modifications to the strain field to alleviate geometric and material locking phenomena [9]. In the context of forming simulations, solid-shell elements have been successfully applied to sheet metal forming [10,11], packaging simulation of cardboard [12] and recently composite forming [13]. Thereby, the prismatic solid-shell element developed by Xiong et al. [13] for thermoforming of thermoplastic prepregs utilizes an additional degree of freedom at the element centre for an improved calculation of transverse normal stresses, in combination with a discrete Kirchhoff assumption for zero transverse shear strains.
Additionally, engineering textiles have a specific forming behaviour, which requires a membrane-bending decoupling to describe their deformation behaviour. Three-dimensional element formulations based on the classical continuum mechanics of Cauchy are not capable of modelling very low stiffness under bending, while simultaneously considering a high in-plane membrane stiffness. Therefore, methods based on generalized continuum mechanics can be utilized. Those so-called second-order gradient approaches require elements with high-order interpolation functions, which are numerically expensive and difficult to implement in a commercial solver [14]. As an alternative, approaches introducing a bending stiffness based on a curvature calculation with neighbouring elements have shown promising results for thick 3D woven reinforcements [15].
In the present study, the 3D hexahedral solid-shell element with only translational degrees of freedom initially proposed by Schwarze and Reese [2,3] and extended to an explicit formulation by Pagani et. al. [16] is used. The selected element formulation is implemented as user-element (VUEL) in ABAQUS/Explicit. To ensure computational efficiency, a MATHEMATICA-based programming environment called ACEGEN [17] is used. ACEGEN provides a symbolic implementation and differentiation combined with a simultaneous runtime-optimization. This work presents a continuation of the investigations performed for the isotropic case in Schäfer et al. [1]. It extends the approach for the forming of highly anisotropic materials and introduces modifications to the hourglass stabilization to account for the anisotropy in the stabilization. The potential of this solid-shell element for the macroscopic forming simulation of engineering textiles is shown by comparison to commercially available shell and solid elements during multiple hemisphere forming tests. Furthermore, an approach for the membrane-bending decoupling within the linear 3D element formulation is proposed. It is based on a Taylor approximation of the strain with respect to the out-of-plane direction and does not require to consider the displacements of neighbouring elements. The potential of this decoupling method is shown by application to numerical coupon and component forming tests.
Solid-shell element
The investigated solid-shell element is derived from a standard isoparametric 8-node hexahedral brick-element with tri-linear shape functions. The details of the element formulation can be found in Schwarze & Reese [2,3], as well as the extensions necessary for an explicit formulation in Pagani et. al. [16]. This section only highlights some of the main aspects of the solid-shell element relevant for a geometric and volumetric locking-free bending behaviour, which is necessary to accurately and efficiently describe the forming behaviour of thin structures with high aspect ratios [1].
The element uses a reduced integration scheme with a single integration point in the shell plane and a variable number of integration points (at least 2) along the normal through the element centre defined in isoparametric coordinates by * = {0, 0, } ⊤ . This integration scheme was shown to be computationally efficient [16] and captures non-linearities in the thickness direction for thin structures very well, compared to commercially available explicit reduced-integration elements which mostly use a single integration point in the element's centre.
The total Green-Lagrange strain tensor is additively split ( = c + e ) according to the enhanced assumed strain (EAS) concept based on the Hu-Washizu variational principle into a compatible part c ( ), depending solely on the displacement vector , and an enhanced part e ( e ), depending on a single additional enhanced degree of freedom e , which is utilized to prevent volumetric and Poisson thickness locking. The transverse shear and curvature thickness locking are cured through the assumed natural strain (ANS) method for the transverse shear and normal components of the covariant compatible Green-Lagrange strain.
Another key aspect of the element is a Taylor approximation of the compatible strain with respect to element centre = : ̂c ≈̂0 +̂ζ + 2̂ζ ζ ⏟ ̂c * +̂c ξ +̂c η +̂c ξη +̂c ηζ +̂c where the derivatives ̂( •) are constant tensors. This allows for a separation into a strain part related to the out-of-plane integration ̂c * and a part unused by the numerical integration ̂c hg , which is essential to the hourglass stabilization of the reduced integration scheme. Similarly, a Taylor approximation of the second Piola-Kirchhoff stress along the out-of-plane direction * is carried out: where again the stress is separated into the parts related to the out-of-plane integration and hourglass stabilization. The complexity of the hourglass stabilization part is reduced by replacing the in general material-and deformation-dependent tangent ̂ ̂| = * with a constant hourglass stiffness matrix Ĉ hg . This method allows for efficient analytical integration of the hourglassing terms while maintaining a full rank of the element formulation. For the description of the material behaviour, a Saint Venant-Kirchhoff hyperelastic approach [18] is used, which leads to ̂ * = Ĉ * ⋅ (̂c * +̂e * ) and ̂h g = Ĉ hg ⋅̂c hg .
Adaptions for forming simulations of engineering textiles The above described solid-shell element has shown very promising results compared to commercially available elements in ABAQUS/Explicit for the forming simulation of thin structures with high aspect ratios in an isotropic case [1]. In the next step, the formulation is tested for highly anisotropic materials, to emulate the deformation behaviour of engineering textiles. Therefore, two different orthotropic stiffness matrices with high stiffness in one (Ĉ UD ) or two (Ĉ Biax ) principal material directions and low shear and compaction stiffnesses for both cases are chosen for this investigation. A relatively low stiffness of 1000 MPa in the principal material directions is chosen to reduce the explicit time increment in this comparative study, while maintaining a large anisotropy ratio and therefore limiting the in-plane strains in fibre direction: The application to model a single-layer hemisphere forming test leads to hourglassing instabilities on the edges during the initial contact of the fabric and tool, cf. Figure 1. The location of these effects agrees with observations from single element tests with as few Dirichlet boundary conditions as possible, where the most critical deformation for provoking instabilities is compaction in the thickness direction. Modified hourglass stabilization. In the work of Schwarze & Reese [2,3], the hourglass stiffness matrix was approximated by a constant deviatoric matrix Ĉ hg = eff hg ̂d ev based on an effective shear modulus eff hg and the deviatoric part of the fourth-order identity tensor ̂d ev . A deviatoric hourglassing stiffness was chosen to ensure a volumetric locking-free stabilization, according to the B-Bar method developed by Hughes [19]. This approximation leads to an underestimation of the hourglassing stress for highly anisotropic materials because the calculated isotropic effective shear modulus is significantly too low. Therefore, the hourglassing stiffness matrix is replaced within this study by the material tangent evaluated in the element centre through which accounts for the anisotropy in the stabilization, while still allowing for an analytical integration of the hourglassing terms. However, this approach does not conform with the B-bar method and is therefore limited to compressible materials to prevent volumetric locking.
The construction of a suitable anisotropic hourglass stiffness, which prevents volumetric locking, would require further investigations and is not part of this work, because the intuitive approach to only use the deviatoric part of the anisotropic stiffness Ĉ hg dev = ̂d ev : Ĉ hg could lead to non-symmetric stiffness matrices and therefore potentially resulting in non-physical negative hourglassing energies. However, since the focus of this work are engineering textiles, which are compressible in most cases, and the choice for the hourglassing stiffness according to Equation 5 shows promising results, the influence of a possible volumetric locking in the hourglassing terms is assumed to be small.
Membrane-bending decoupling.
To describe the forming behaviour of engineering textiles, a decoupling of the in-plane (membrane) and out-of-plane (bending) behaviour is often introduced in approaches based on conventional shell elements. Based on this idea and in combination with the Taylor approximations of the strains in Equations 1, the following split is introduced to the stress ̂ * by ̂ * =̂m em +̂b end with ̂m em = Ĉ mem ⋅̂c 0 and ̂b end = Ĉ bend ⋅ (̂c + 1 2 2̂c +̂e * ), where the membrane part ̂m em is related to the constant compatible strain in the element centre ̂c 0 by a membrane stiffness matrix Ĉ mem , while the bending part ̂b end is related to the out-of-plane components of the compatible strain ̂c and ̂c as well as the enhanced strains ̂e * by a bending stiffness matrix Ĉ bend .
Numerical studies
To investigate the potential of the solid-shell element formulation outlined in Section 2 for forming simulations of engineering textiles, two different numerical studies are shown in the following section. First, the advantages of a geometric and volumetric locking-free forming behaviour of the solid-shell element for highly anisotropic materials are presented through a comparison to in ABAQUS/Explicit commercially available shell and solid elements in a single layer hemisphere test. Similar to our previous study [1], where the significant influence of the element formulation on the forming behaviour even for a simple linear isotropic material model was shown. Secondly, it is demonstrated by a simple tensile and cantilever-bending test that the approach proposed in Equation 6 can be utilized for a membrane-bending-decoupling, and initial results for its influence on the forming behaviour are shown.
Hemisphere test for highly anisotropic material
To focus on the influence of the element formulation on the forming behaviour as best as possible, a single layer with a thickness of = 0.3 mm is used in multiple hemisphere tests with a total tool stroke of 35 mm, cf. Figure 1 a). The tests are performed with an orthotropic material law, for both stiffness matrices Ĉ UD and Ĉ Biax , cf. Equation 4, as well as aspect ratios ( = e / ) of 1 = 10 and 2 = 20. The solid-shell element without the membrane-bending decoupling is compared to in ABAQUS/Explicit commercially available reduced integrated shell (S4R) and solid (C3D8R) elements, as well as a fully integrated solid element (C3D8), a reduced integrated solid element with an enhanced hourglass stabilization based on the EAS method (C3D8R-Enh) and a fully integrated element with incompatible modes (C3D8I). Figure 2 shows the results for a remaining tool stroke of Δ = 7.0 mm because at this stage the differences in the forming behaviour can be better highlighted than for a completely closed tool. The wrinkling development in terms of size and pattern is strongly influenced by the different element types, whereas the elements known to behave too stiff under bending deformation tend to have larger wrinkles form. These results agree with the observations made in literature [20][21][22][23].
Based on the assumption that the reduced integrated shell element (S4R) is locking-free even for large the aspect ratios [1,24], it is used as a desired reference solution in this study. For a biaxial stiffness Ĉ Biax and both 1 as well as 2 , it shows a consistent forming behaviour with medium-sized wrinkles developed in all directions. For a unidirectional stiffness Ĉ UD the results are again similar for both investigated aspect ratios, with the developed wrinkles mainly oriented along the axis of the highest stiffness and semi-circular runouts towards the edges. The only element that can achieve a forming behaviour similar to the shell for all tests is the solid-shell element presented in Section 2. However, looking into more detail some individual elements in the area of radius shows a slight hourglassing shortly before the tools are completely close even though it seems not to influence the global forming behaviour. This indicates that the proposed approach for the hourglass stiffness, cf. Equation 5, still needs improvements in areas of large compactions.
The deformation behaviour of the reduced integrated solid element (C3D8R) is completely different compared to all other examined elements. The orientation of the onset of wrinkles is only perpendicular to the edges of the thin sheet and they are larger compared to the reference solution. While introducing an enhanced hourglassing stabilization (C3D8R-Enh) to alleviate lockingeffects showed very promising results in the previous study for isotropic material [1], the behaviour for highly anisotropic materials is significantly different with an unexpected orientation of the development of wrinkles in the y-direction for a unidirectional (highest stiffness in the x-direction) as well as a biaxial stiffness. The results of the fully integrated solid element (C3D8) in the biaxial case for a smaller aspect ratio 1 are in good agreement with the chosen reference. An increase in size and decrease in the number of wrinkles is observed for the larger aspect ratio 2 , which could also be seen for the isotropic case and is expected due to its known too stiff behaviour in bending situations due to locking phenomena [1]. However, in the unidirectional case, the orientation of the onset of wrinkles is slightly different and the size is increased, which again indicates a too high bending stiffness. The fully integrated solid element C3D8I is enhanced by incompatible modes to prevent parasitic locking stresses in bending deformations [22]. Thus, it shows a forming behaviour very similar to the shell element for a smaller aspect ratio 1 and only slightly too stiff for 2 .
The hemisphere tests show that the element formulation and thus locking-free bending behaviour for larger aspect ratios have a significant influence on the macroscopic forming behaviour, even for a simple linear orthotropic material model. The solid-shell element, cf. Section 2, is the only element which consistently has a similar forming behaviour compared to the chosen reference shell element (S4R) for large aspect and anisotropy ratios.
Membrane-bending decoupling
Proof of concept. To test whether the membrane-bending decoupling based on the approach proposed in Equation 6 works, simple tensile and cantilever-bending tests are performed. The geometry and load conditions are shown in Figure 3. In both tests, the beam is discretized by one element in the thickness and 10 elements in the length direction, which yields an aspect ratio of = 20. Both tests are performed with a total of four different combinations of a higher and lower membrane as well as bending stiffness each. The results of the tensile and cantilever-bending tests are shown in Table 2 (a) and (b) respectively. As expected, the tip displacement in the tensile test is only increased by decreasing the membrane stiffness and independent of the chosen bending stiffness. Analogously, this behaviour is shown for the cantilever-bending test with a decrease of the bending stiffness. This indicates that the proposed approach for a membrane-bending-decoupling in Equation 6 works. Influence on the forming behaviour. The proposed membrane-bending decoupling is applied to single-layer hemisphere tests, cf. In the biaxial case, decreasing the bending stiffness leads to the onset of smaller wrinkles and an increase in their number, while their general orientation, as well as location, stays the same. A higher bending stiffness in the unidirectional case leads to a straightening of the wrinkles located towards the edges, while for a lower bending stiffness more wrinkles perpendicular to the xaxis are forming. These results agree with the observations made in the literature [20][21][22][23].
Conclusion and outlook
A linear solid-shell element suitable for the forming simulations of highly anisotropic materials with a membrane-bending decoupling based on a Tailor expansion of the strain is presented. The advantages of its locking-free bending behaviour in forming simulations are shown by comparison to commercially available shell and solid elements. The forming results are consistent for large aspect ratios and in agreement with a typical shell element, outperforming all investigated commercially available solid elements that either behave significantly different or too stiff due to locking phenomena. The proposed approach can show typical characteristics of the membrane-bending interdependency in coupon tests as well as in forming simulations. The shape, size and orientation of wrinkles depend on the ratio between the membrane and bending stiffness.
These are necessary prerequisites to model the deformation of thin engineering textiles efficiently in macroscopic forming simulations. However, so far only linear-elastic orthotropic materials are investigated. In future studies, the material model will be adapted to the in general non-linear behaviour of different deformation modes of an engineering textile and validated with experimental test. In this context, especially the proposed calculation of the hourglassing stiffness needs to be re-evaluated, since during the numerical studies some slight hourglassing occurred in areas of higher compaction.
The utilized solid-shell element formulation is known to be geometric and volumetric locking-free for isotropic and slightly anisotropic materials [1][2][3]. However, the potential occurrence of a locking phenomenon specific to engineering textiles known as intra-ply shear [25] or tension [26] locking needs to be investigated. In this study, the impact of this locking phenomenon is assumed to be small, since the principal material directions are aligned with the mesh orientation. This alignment is not always possible for arbitrary shaped textiles and may require further modifications to the hourglass stabilization to alleviate tension locking [26].
Furthermore, the benefits of modelling the thickness direction will be investigated in the context of composite forming. Since the state of the art forming approaches are mainly based on shell elements and therefore often limited in the consideration of the out-of-plane behaviour. The utilization of a fully three-dimensional approach should enable a better approximation of forming effects like local thickness changes and fibre volume content [6], by investigating the compaction behaviour. | 4,835.6 | 2021-04-14T00:00:00.000 | [
"Materials Science"
] |
Mathematical modelling of impurity deposition during evaporation of dirty liquid in a porous material
Abstract When a contaminated liquid evaporates from within a porous material, the impurities or dirt accumulate and deposit within the pore space. This occurs during the cleaning of filters and fouling of textiles, and is related to the ‘coffee-ring’ problem. To investigate how and where dirt is deposited in the pore space, we present a model for the motion of an evaporation front through a porous material, and the related accumulation, transport, and deposition of dirt, assuming that the liquid remains stationary. For physically relevant parameters, vapour transport out of the porous material is quasi-steady and we derive a single ordinary differential equation describing the motion of the evaporation front in time. Model solutions exhibit spatially non-uniform profiles of the deposited dirt-layer thickness through the porous material. The dirt accumulation and evaporation problems are coupled: deposited dirt hinders vapour transport through the porous material, slowing the evaporation. We identify two scenarios in which the porous material becomes clogged with dirt. Accumulation of suspended dirt at the evaporating interface along with slow dirt diffusion results in the deposited dirt layers clogging the pores at the evaporating interface, halting the drying and trapping liquid in the porous material. Alternatively, slow dirt deposition results in the suspended dirt being pushed far into the porous material by the evaporation, eventually leaving only dirt (with no liquid) in the pore space. We investigate the dynamics of both clogging scenarios, characterising the parameter regimes for which each occurs. Both clogging scenarios must be avoided in practice since they may be detrimental to future filter efficacy or textile breathability.
Introduction
Drying-driven redistribution of dirt within filters and textiles is a common problem, with practical industrial importance.For instance, after the rinsing of filters used in vacuum cleaners or washing machines, the filter dries and any remaining dirt or cloth fibres are left in the filter (Ji & Sanaei 2023), reducing its capacity for the next filtration cycle.Waterproof clothing such as coats and boots will dry after use, and impurities or dirt may similarly be left within the pores of the textile and waterproof membrane (Breward et al. 2020;Sanaei et al. 2022).
A key question in these filtration and waterproof-clothing applications is to determine where within the porous material the dirt is deposited once the liquid has all evaporated.We might also ask whether all of the liquid can indeed be evaporated, or whether it becomes trapped in regions of pore space clogged by the deposited dirt.In the applications of interest, it is important that the dirt does not clog the material, as this leads to reduced filter efficacy, or reduced breathability of the waterproof garment.Furthermore, trapped water in a washing machine filter may contribute to the growth of bacteria or mould, and should be avoided for hygiene reasons (Abney et al. 2021).A paradigm situation encompassing these processes is that of a porous material containing a mixture of a liquid, such as water, and an impurity or dirt that is suspended in the liquid.As the liquid evaporates, an evaporation front moves into the porous material from its surface.The dirt is left behind in the liquid as the liquid evaporates, and may deposit into a layer on the walls of the pore space.
A related problem is the deposition of suspended particles when a droplet of liquid dries on an impermeable substrate.This is known to lead to a coffee-ring effect, in which the coffee particles are transported by evaporation-induced flow of liquid to the edge of the droplet.This coffee-ring effect is well studied, for instance, by Deegan et al. (1997Deegan et al. ( , 2000)), Karapetsas, Sahu & Matar (2016), Kaplan & Mahadevan (2015), Moore, Vella & Oliver (2021), Murisic & Kondic (2011), Popov (2005), and recently reviewed by Wilson & D'Ambrosio (2023).The coffee-ring effect is of practical importance, for instance, in the drying of ink droplets in ink-jet printing (Mampallil & Eral 2018;Soltman & Subramanian 2008) and in the manufacture of electronic devices (D'Ambrosio et al. 2021).
In a drying porous material, such as a filter membrane or textile, there are several additional complications not present in the coffee-ring set-up.Firstly, the problem is multiscale in nature, in that the fluid flow, evaporation and the transport and deposition of the dirt occur within the pore space, while the depth of porous material to be dried is likely to be significantly larger than an individual pore, even for fairly thin filter membranes.It is not immediately clear how to formulate a model that captures the pore-scale behaviour and yet remains tractable over the scale of the entire drying material.Additionally, dirt deposition may occur throughout the porous domain, not only at the base of the evaporating droplet.This means that there is additional coupling between the drying and the deposition: like in evaporating droplets, the accumulation of suspended dirt at the evaporating interface can reduce the evaporation rate (Karapetsas et al. 2016), but additionally the build-up of deposited dirt in the pore space affects the porosity and reduces the rate of diffusive transport of (i) the suspended dirt through the liquid-saturated pore space, and (ii) vapour out through the dry porous material.In extreme cases, the deposited dirt might completely clog the pore space at the evaporation front, terminating the drying before all liquid is evaporated.This phenomenon is not possible in coffee-ring problems.Like the surface tension driven flows in coffee rings, a capillary flow may draw fluid through the porous material, which then evaporates near the surface of the porous material (Lehmann, Assouline & Or 2008).This is typically the case early in the drying process ('stage I'), while later ('stage II') the evaporating interface moves into the porous material, and the transport of vapour out of the porous material limits the evaporation rate (Or et al. 2013;Fei et al. 2022).
Drying porous media (without dirt) have been studied in a variety of settings and using various different modelling techniques.Depending on the porous material and fluids, various drying regimes are possible: liquid and gas may coexist within the pore space throughout the entire medium and for the majority of the drying time (so that the majority of the drying is in stage I), or a region in which capillary effects dominate, often referred to as a 'film region', may separate a region of porous material saturated with liquid from a multiphase region incorporating unconnected pockets of stationary liquid (Pel, Landman & Kaasschieter 2002;Lehmann et al. 2008).Multiphase flow models for drying are derived by, for instance, Whitaker (1977) while lumped models, consisting of nonlinear diffusion equations for the 'moisture' (combining liquid and vapour) are also often used (Vu & Tsotsas 2018;Pel et al. 2002).Evaporation within the pore space may be simulated directly, although this is computationally expensive and limited to sufficiently small domain sizes (Fei et al. 2022).Pore-network models are a more computationally tractable approximation, although the details of the fluid flow are neglected (Nowicki, Davis & Scriven 1992;Tsimpanogiannis et al. 1999).
The transport and trapping of particles in a liquid-saturated porous material when the liquid is flowing is known as deep-bed filtration (Zamani & Maini 2009).When there is no flow of the liquid, the particles may still be transported by Brownian diffusion (Epstein 1988).Particles may build up in a deposited layer on the pore walls due to several mechanisms, including electrostatic forces in the bulk (Zamani & Maini 2009).Particles are generally repelled from air-water interfaces unless they are hydrophobic; in the hydrophobic case they might be held at the interface and, thus, transported more effectively with it (Flury & Aramrak 2017).Particles may deposit or attach to the walls of the pore space due to adsorption, electrostatic forces or other chemical binding mechanisms (Epstein 1988;Zamani & Maini 2009;Dressaire & Sauret 2017).Experimental work such as that of Gudipaty et al. (2011), Linkhorst et al. (2016), Stamm et al. (2011) seeks to visualise the deposits and quantify their growth rates in terms of the system parameters.
For an evaporating droplet, an evaporative flux is generally prescribed at the droplet surface (Popov 2005;Murisic & Kondic 2011;Kaplan & Mahadevan 2015;Karapetsas et al. 2016;Moore et al. 2021).This flux may be constant (Moore et al. 2021), but typically depends on the distance from the edge of the droplet, accounting for the quasi-steady transport of vapour away from the droplet (Popov 2005;Karapetsas et al. 2016).When drying from within porous media, a prescribed evaporation rate may be appropriate during stage I (when the evaporation occurs near the surface of the material) but, since the stage II drying of porous media is limited by the removal of vapour from the pore space (Lehmann et al. 2008), like the majority of evaporating drops (Wilson & D'Ambrosio 2023), we expect that the evaporation rate will depend on the position of the evaporating interface within the porous material during this stage.
The deposition of dirt during the drying of a filter has recently been studied by Ji & Sanaei (2023).Here, the suspended dirt is assumed to diffuse through a liquid-saturated region of porous material ahead of an evaporating interface, and deposit at a rate directly proportional to its concentration, causing the local porosity to decrease.The evaporating interface is assumed to move through the porous material at a prescribed speed, dependent only on the local porosity and suspended dirt concentration, and not the location of the front within the filter.Simulations of this model show that the porosity of the filter decreases during the drying, and that the deposited dirt is non-uniformly distributed in the pore space once the drying is complete.
In this paper we systematically derive a homogenised model for the coupled processes of evaporation, transport of liquid vapour, diffusion and deposition of dirt in a drying porous material, starting from a pore-scale model for these processes.This analysis is based on previous work (Luckins et al. 2023) for evaporation of a pure liquid in a porous material, extended to incorporate dirt transport and deposition.One benefit of the homogenisation approach is that the pore-scale behaviour is included in the homogenised model through averaged terms.This ensures that the model conserves mass of all species, and also results in a different diffusive term in the homogenised equations compared with the model posed by Ji & Sanaei (2023).For simplicity, as in both Luckins et al. (2023) and Ji & Sanaei (2023), we assume that capillary flows are negligible and a sharp evaporating interface moves into the porous material.In practice, such systems would be valid when viscous or gravitational forces dominate over surface tension, for instance, if the solid is hydrophobic (Shokri, Lehmann & Or 2008) or the pores are sufficiently large relative to the capillary length scale (Lehmann et al. 2008).Our coupled model for the drying and dirt transport is a type of Stefan problem, with undercooling in certain parameter regimes.We derive our homogenised model in § 2. In § 3 we note that the vapour transport is quasi-steady for physically relevant parameter choices and reduce the vapour-transport problem to a single ordinary differential equation (ODE) for the position of the evaporation front, providing a comparison between this model and that of Ji & Sanaei (2023).In § 4 we study the early time behaviour of our model and describe our numerical solution method.In § § 5-6 we study the asymptotic limits of slow and fast deposition rates, identifying a distinct mechanism in each case by which the system may clog before the drying is complete.We quantify the parameter regimes for which these clogging phenomena occur in § 7 before concluding in § 8.
Model derivation
We consider a porous material of finite thickness l, initially with uniform porosity and saturated with a uniform mixture of liquid and suspended dirt.We assume that the dirt particles are small relative to the pore-length scale, and neither interact with each other nor dissolve in the liquid.The dirt-liquid mixture is thus a suspension of these insoluble dirt particles.We suppose the porous material is bounded by an impermeable solid material on one side.The liquid begins to evaporate from the side open to the atmosphere, leaving the dirt behind, and an evaporation front moves into the porous material, with the suspended dirt and liquid ahead of the front, and a mixture of inert gas (drawn in from above the porous material) and liquid vapour behind it.We assume the system is isothermal, with no variation in temperature.A schematic of the situation under consideration is shown in figure 1.We consider a two-dimensional porous material for simplicity, with spatial variables x and y, and with y pointing into the porous material and y = 0 at the surface of the porous material.Although the structure of our model and the homogenisation analysis does not depend on the pore-scale geometry, it is helpful to specify this for simplicity.We choose a square lattice of circular solid inclusions, of radius r 0 .We account for deposition of the suspended dirt onto the solid structure by considering deposited dirt layers of thickness R(x, y, t), on each solid inclusion, which have initial thickness zero.An important assumption is that the liquid-dirt mixture does not flow, and so our model excludes any capillary pressure or surface tension effects (since in order to attain a (quasi-)static meniscus shape, the liquid would need to flow).We first consider the drying behaviour on the microscale -within the pores of the material -before averaging to derive our effective model.We suppose that the evaporating front is located at y = h(x, t), splitting the domain into a region of pore space containing vapour in 0 ≤ y ≤ h(x, t), where the thickness of the layers of deposited dirt do not change with time, and a region of pore space in h(x, t) ≤ y ≤ l containing the liquid-dirt mixture, where the dirt-layer thicknesses vary in time due to deposition or erosion.
Pore-scale model
In the pore space occupied by the vapour-gas mixture (behind the evaporating front, i.e. y < h(x, t)), we expect the Reynolds number to be small (Luckins et al. 2023) and so we assume that the mixture satisfies the Stokes equations where u is the mass-averaged velocity of the mixture, p is the pressure and μ is the viscosity (assumed constant).The vapour contained within the mixture is advected with the flow, and also diffuses through the mixture with diffusivity D v , and thus, the density of the vapour, ρ v [kg m −3 ], satisfies where the subscript t denotes partial derivative.The overall density of the inert gas-vapour mixture, ρ G , is given by (2.3) Wherever the gas-vapour mixture meets the solid walls of the pore space we suppose there is no flux or slip of the gas-vapour mixture, and no flux vapour into the solid material, so that on the solid-liquid or dirt-liquid boundary, with normal (2.4a,b) In the liquid-dirt mixture in y > h(x, t), we assume that there is no net flow of the mixture, and the suspended dirt and liquid diffuse against one another in an ideal mixture, due to Brownian motion.The suspended dirt volume fraction, θ, therefore satisfies (2.5) where D d is the diffusivity of suspended dirt in liquid.As discussed in Luckins et al. (2023), the assumption that the liquid does not flow means that capillary effects are neglected from the model.
At the solid walls of the pore space, we suppose that the suspended dirt can deposit onto the solid microstructure, forming a layer that may also then be eroded away.We suppose that the deposited layer has a dirt volume fraction θ * which is the packing volume fraction of the dirt particles.We expect this to be close to one, as only a small amount of liquid (volume fraction 1 − θ * ) is trapped within the deposited dirt layer.Conservation of dirt across the interface is given by θ (2.6) where V n is the normal velocity of the depositing/eroding interface.We note that, in order that there is no flow generated at the depositing interface, we assume that the dirt and liquid have the same mass density, so that the total mixture density is the same on either side of the depositing/eroding interface, while the dirt and liquid fractions can jump (see, for instance, Geng, Kamilova & Luckins 2023).
We suppose that the dirt is deposited at a rate dependent on the local suspended dirt volume fraction, while the erosion rate depends on the (constant) packing volume fraction θ * .(If there was a flow of the fluid, we might extend this model and allow the erosion rate to depend on the local shear stress.)Thus, we prescribe (2.7) where the constants k ± have units m s −1 .This type of law-of-mass-action deposition rate, in which the deposition rate is linear in the quantity of suspended dirt, is common in the phenomenological bed-filtration literature (Zamani & Maini 2009;Dressaire & Sauret 2017), and is also used by Ji & Sanaei (2023) as a model for adsorption of particles onto the deposit layer.
At the evaporating interface y = h(x, t), we suppose that the inert gas and the dirt do not pass through the interface, while liquid turns into vapour.We thus impose conservation of mass of each of the liquid/vapour, gas, and suspended dirt, namely (2.8c) where ρ l and ρ d are the densities of pure liquid and dirt, respectively.Combining these, we derive the more helpful form interpretable as a condition on each of u • m, ρ v and θ, respectively, where ρ L = ρ l (1 − θ) + ρ d θ is the (assumed constant) liquid-dirt mixture density.The normal velocity of the interface and unit normal to the interface are given by where subscripts t and x denote partial derivatives.In addition to (2.9), we also impose a no-slip condition for the gas-mixture velocity u + vh x = 0 on y = h(x, t). (2.11) Finally, we must also incorporate a condition that describes the chemistry governing the evaporation.In Luckins et al. (2023) the effect of different chemistry conditions were considered, and these were shown to affect the form of macroscale boundary conditions derived through a homogenisation analysis.For simplicity, we assume that the liquid and vapour are in chemical equilibrium at the evaporating interface.The chemical potential on the liquid side of the interface is dependent on the amount of liquid (1 − θ) at the interface, while the chemical potential on the gas-mixture side depends on the density of vapour, ρ v , at the interface.In general, we may express this chemical equilibrium as (2.12) where ρ * is the (constant) saturation vapour density when θ = 0 and there is no suspended dirt, and the function f (θ ) captures the dirt dependence of the saturation vapour density.
(We note that therefore f (0) = 1.)The presence of particles at the interface are expected to hinder the vaporisation; in both Ji & Sanaei (2023) and Karapetsas et al. (2016) the evaporative flux is modelled as decreasing with increased particles on the fluid surface.We keep f (θ ) general as far as possible, and in § 5 we investigate the effect of different functional dependencies f (θ ) on the drying rate.However, in our numerical simulations we use the simple linear form to capture the effect of the dirt inhibiting vaporisation.We choose this form so that the saturation vapour density scales with the amount of liquid at the interface.At the surface of the porous material, we impose a constant atmospheric vapour density and atmospheric pressure ρ v = ρ a , p = p a , on y = 0. (2.14) Dirt cannot diffuse through the impermeable boundary and thus so impose that This depth l is assumed to be much greater than the typical pore-length scale, so that there is separation between the pore-and macro-length scales.
Non-dimensionalisation
We non-dimensionalise the vapour/gas problem in a similar way to Luckins et al. (2023), making the rescalings are dimensionless parameters representing the ratio of vapour density to liquid density, the ratio of pore-to macro-length scales and the ratio of vapour to gas densities, respectively.
In particular, we note that we have chosen the time scale associated with the speed of the motion of the evaporating interfaces on the microscale, i.e. the time scale over which sufficient vapour is removed by diffusion to empty the microscale pore space of liquid.Making these rescalings and dropping the hat notation, the dimensionless microscale model is, in y < h(x, t), (2.18b) At gas-solid interfaces in y < h(x, t) (which are stationary), and at liquid-solid interfaces in y > h(x, t) (which move with velocity V n ), while at the evaporating interfaces y = h(x, t), At the surface of the porous material, while at the impermeable surface (or centre of a symmetric porous material), Here we have introduced the additional dimensionless parameters that appear in the dirt problem, representing the ratio of the suspended dirt diffusion time scale to the time scale of the evaporation-front motion, the ratio of the dirt-deposition rate to the evaporation rate, the ratio of the dirt erosion rate to deposition rate and the ratio of the atmospheric vapour density to the maximum saturation vapour density, respectively.
The micro-to macro-length scale ratio (defined in (2.17a-f )) is the small parameter we take advantage of in order to homogenise (2.18).As in Luckins et al. (2023), we take δ < 1 and ν < 1 to be order one parameters relative to for the homogenisation analysis.We note that δ ≈ 10 −3 for water, so will later consider the additional limit of δ → 0 (which is equivalent to taking this limit before performing the homogenisation analysis).We require α < 1 but expect α 1 to be reasonable.Although the diffusion of vapour in air is generally much faster than the diffusion of any kind of molecule through a liquid, so that D v D d , we note that since δ ≈ 10 −3 -10 −4 is small, σ is likely to be order one.For instance, if we take D v ≈ 2.5 × 10 −5 m 2 s −1 and D d ≈ 10 −9 m 2 s −1 (Cussler 2009), then with δ = 10 −4 we find that σ ≈ 2.5.We consider the distinguished limit of σ = O(1) in this paper, but we note there is an alternative, slow-dirt-diffusion, distinguished limit with σ = O( −1 ).We discuss this alternative case further in Appendix A.2, and briefly in § 2.3 below.In summary, all of σ, κ, β and α are taken to be order one relative to to homogenise the model.
Summary of the homogenised drying model
The homogenisation analysis is described in Appendix A. The result of this analysis is a macroscale model for the vapour density ρ, suspended dirt volume fraction θ, deposited dirt-layer thickness, R, and position, Y = H(T) of the evaporation front, namely for Y < H(T), and and effective diffusivity, D(R) (given by (A4)), all vary with the thickness of the deposited dirt layer.We assume that, initially, the porous material is entirely saturated with a uniform liquid-dirt mixture, none of which has yet deposited (i.e. the time scale of deposition is assumed longer than the time scale over which the liquid-dirt mixture flooded the material).Thus, at T = 0, (2.21a-c) Our homogenised model (2.20) is similar in structure to those proposed in Breward et al. (2020), Sanaei et al. (2022) and Ji & Sanaei (2023), with the suspended dirt satisfying a reaction-diffusion equation ahead of a moving evaporation front.However, through the systematic homogenisation analysis, we have found the correct form for the diffusion term in (2.20b), which was erroneously given as (D(φθ ) Y ) Y (in our notation) by Ji & Sanaei (2023).Additionally, we have quantified the effective parameters D, C and φ, which all vary with the deposited dirt-layer thickness R. We also impose different boundary conditions to Ji & Sanaei (2023), which will result in different drying behaviours, as discussed in § 3 below.
A key assumption of our homogenisation analysis was that σ = O(1), which ensured that θ (and therefore R) is uniform to leading order on the microscale.As discussed further in Appendix A.2, the extremely slow suspended dirt diffusion limit of σ = O( −1 ) is not captured by this model: in this case we would expect non-periodic behaviour on the microscale at the evaporating interface, and the homogenisation analysis would break down.We do not consider this situation here.
An ODE for the motion of the evaporation front
The parameter δ = ρ * /ρ L is generally small; indeed for water evaporating we expect δ ≈ 10 −3 .Before studying the full drying problem, we consider the limit of δ → 0 in the vapour-gas transport problem, which we show results in a single ODE describing the motion of the evaporation front H(T).This gives insight into the drying dynamics and is interesting as a comparison with other models for the motion of evaporating interfaces in the literature, e.g.Ji & Sanaei (2023).Furthermore, the analysis in this section is helpful for all of the subsequent analysis of the model, including the early time asymptotic analysis in the following section ( § 4), which we use to initialise numerical simulations of the model.
For small δ, we see from (2.20a) that the vapour-density profile is quasi-steady, varying instantaneously with the motion of the evaporation front.Specifically, in the limit δ → 0, the vapour-transport equation (2.20a) becomes (3.1) Integrating twice with respect to Y and applying the boundary conditions (2.20d) and (2.20g), we obtain By additionally imposing the boundary condition (2.20e) we obtain an equation for the motion of the evaporation front, H, in terms of the suspended dirt volume fraction there, θ| H , namely One particular case of interest is if D is uniform (for instance, if little dirt has been deposited, so R ≈ 0 is constant).In this case (3.3) reduces to If θ| H were constant, we would see a √ T behaviour of the evaporation front, as expected for this type of Stefan problem.For D non-uniform in Y, the integral term in (3.3) behaves like an overall resistance to vapour transport.In particular, the integral is dominated by any localised regions of pore space in Y < H for which D is very small.We see that the ODE (3.4) for H (with D constant) takes the form for an algebraic function E 1 , while the more general (3.3) takes the form By comparison, Ji & Sanaei (2023) prescribe an evaporative flux that does not explicitly depend on H, of the form in our notation.Unlike (3.5) and (3.6), the model (3.7) does not explicitly depend on the position H of the evaporating interface.These different equations for H result from different modelling assumptions: Ji & Sanaei (2023) assume that the vaporisation of the liquid molecules is the limiting process in the evaporation, whereas we have assumed that the vaporisation is instantaneous (the vapour is at its saturation point adjacent to Y = H) and that evaporation is instead limited by the transport of vapour out of the porous material.
For sufficiently deep or hydrophobic porous media that there is a moving drying front, it is clear that the evaporation rate should depend on the location H of the drying front (Lehmann et al. 2008;Shokri et al. 2008).Furthermore, we note that our drying model is given in terms of well-defined physical parameters such as the diffusivity and saturation vapour densities, and results in a reasonable drying time scale l 2 ρ L /ρ * D v ≈ 10 2 s (using values for water: whereas the coefficients in a prescribed evaporation rate must be fitted in some way.We note from (3.3) that evaporation only occurs when f (θ | H ) > α, so that the vapour density at the liquid-gas interface is greater than the atmospheric vapour density; otherwise if f (θ | H ) = α, we see that H T = 0. We define θ such that f ( θ) = α, noting that since f is monotonic in θ, evaporation only occurs for θ < θ.
Our analysis above (and in the remainder of this paper) is for the case that the atmospheric vapour density ρ = α is prescribed at the surface Y = 0.As an aside, we now briefly consider an alternative case, in which the flux, J, of vapour out of the material at Y = 0 is prescribed by a Newton cooling law: J = m(ρ| 0 − a ∞ ), for some constants a ∞ (the far-field ambient vapour density) and m (the mass-transfer coefficient).Since the vapour flux is spatially uniform throughout Y < H in the limit δ 1, we find that φ| H H T = m(ρ| 0 − a ∞ ).Eliminating ρ| 0 , we find that H T is given by the implicit ODE (We may rearrange (3.8) to give H T explicitly in terms of a Lambert-W function, but we consider the form (3.8) more useful as we may compare directly with (3.3).)Clearly in the limit as m → ∞ (for which vapour is easily removed from the surface of the porous material), and with a ∞ = α we regain (3.3).In the case of a bounded mass-transfer coefficient m, the non-instantaneous removal of vapour from the surface results in a slower evaporation rate H T , compared with that given by (3.3).
Early time behaviour and numerical method
In this section we first consider the early time behaviour of our model (2.20) in § 4.1, in the limit of δ
Early time analysis
To study the early time behaviour of the system (2.20), we suppose T = bτ where b 1 is the smallest parameter in the system, and τ = O(1).From (2.20c), on this time scale we see that R τ = O(bκ) 1, and so R is small, hence, all of D, φ and C are constant to leading order in b.
We first consider the vapour problem in Y < H. On the short time scale the interface only moves a short distance, and so we rescale in order to balance the mass-flux boundary condition (2.20d).The vapour problem is therefore self-similar in that we regain the same system at early time with these rescalings as the full system (2.20a), (2.20d)-(2.20e)and (2.20g), namely We have already noted that δ 1 in general, and we take this limit now to make analytical progress.As in § 3, we find that where H(τ ) is the solution of The value of θ at Ȳ = H(τ ) depends on the solution of the suspended dirt problem in the domain Y ∈ (H, 1) = ( bD/φ H(τ ), 1).On this short time scale, the full dirt problem (2.20b), (2.20f ) and (2.20h), with the initial condition (2.21a-c), becomes To leading order in b, we see that θ τ = 0, so that θ = θ IC is independent of time over the domain.However, in a boundary layer at Y = bD/φ H, suspended dirt accumulates due to the motion of the evaporation front.To examine this region, we make the change of variables Y = bD/φ( H(τ ) + z), so that, at leading order in b, the equations are This system (4.6) must be solved with (4.4) to determine θ and H.
We look for a similarity solution of the form for some constant λ to be determined.In particular, from (4.4) we see that the suspended dirt volume fraction at the evaporating interface, θ| H = Θ(0), must be constant in time for such a similarity solution to exist.Substituting into (4.6),we find the solution where λ and the constant Θ(0) satisfy Solutions of (4.9)-(4.10)may be computed numerically, and are shown for various σ and θ IC in figure 2.
To establish some intuitive understanding of this early time behaviour, we now consider the sublimits σ 1, θ IC 1 and σ 1 in turn.We show our early time analytic 986 A31-13 solutions for each case in figure 3 (alongside numerical solutions for comparison, computed using the method described in § 4.2 below), all with excellent agreement.In each, we see that the vapour density ρ varies from f (θ | H ) = 1 − θ| H at Y = H to α = 0 at the surface of the material, according to (4.3).The evaporation front moves with the expected √ T behaviour, faster if there is a steeper vapour-density gradient.Suspended dirt accumulates in the liquid ahead of the evaporation front, with a spatial maximum in θ at Y = H.The size of the boundary layer at H over which θ varies is dependent on σ , which quantifies suspended dirt diffusion.At early times, we expect little dirt deposition, so that the dirt-layer thickness R ≈ 0 throughout the porous material.
If σ 1 so that suspended dirt diffusion is fast relative to the motion of the evaporation front, then we see from (4.9) that λ = O( √ σ ), and so from (4.10) that θ| h ∼ θ IC .
Specifically, we find that Thus, reverting to our original variables, the early time evaporating interface is given by We also note from the form (4.8) of the solution that the spatial region over which θ varies is wide, O(1/ √ σ ).In this small-σ limit, the diffusion of dirt is fast relative to the motion of the evaporation front, and so the suspended dirt volume fraction θ remains close to its initial value θ IC , only deviating by a small, O( √ σ ), amount.For conservation of overall dirt, the region over which the accumulating suspended dirt is spread is wide, of O(1/ √ σ ) relative to the early time boundary layer.This may be seen in figure 3(a,b): since σ 1, the θ profile is approximately uniform in Y, and so close to its initial value of θ IC = 0.1 at early times.The suspended dirt is accumulating due to the evaporation, but spread almost evenly through the domain.
Next, we suppose that σ = O(1) but the initial suspended dirt volume fraction θ IC 1 is small.In this case, for a balance in both of (4.9) and (4.10), we must have Θ(0) = O(θ IC ) and λ = O(1), as we might expect.The solution shown in figure 3(c,d) is for this case, with θ IC = 0.1: we indeed observe that θ| H = O(0.1)(and this effect becomes increasingly clear for smaller θ IC ).
Finally, if σ 1, so that the diffusion of suspended dirt is slow relative to the motion of the evaporation front, then from (4.9) we see that we must have f (Θ(0)) = α to leading order in σ −1 1.At this value, there is no evaporation at leading order, as the vapour density at the atmospheric value is in equilibrium with the liquid-dirt interface, and there is no transport of vapour out of the porous material.Indeed, we see from (4.10) that when θ = θ, λ is the solution of which is independent of σ and of O(1), so that the position of the evaporating interface, given by is of order σ −1/2 1 away from its initial position.Thus, when the suspended dirt diffusion is slow, the diffusion of dirt away from H limits the speed of the evaporation front, so that there is a slower, O(σ ), drying time scale.We also note from (4.8) that the region over which θ varies is narrow, with width O(1/ √ σ ) relative to the early time boundary layer.The boundary layer is narrow for large σ , so that the early time solution actually remains valid for the majority of the drying process.Indeed, in figure 3(e, f ) we see excellent agreement between the early time analytic solution and the numerical solution for ρ, θ and H up to the time when the evaporating front is halfway through the domain.This is because the boundary layer at H over which θ varies is narrow, and so the effect of the boundary at Y = 1 is not felt until H is close to 1.However, we notice in figure 3(e) that the early time approximation R = 0 ceases to be accurate at these late times.The early time solution would remain valid so long as R remains relatively small (e.g. if κ and θ IC are fairly small).We note that, since the boundary layer width scales with √ T, it quickly becomes numerically impractical to resolve the solution at small times for large σ .The early time asymptotic solution is therefore very valuable in initialising the simulations accurately for large σ .
Finally, we note that our early time analysis in this section is equivalent to studying the original problem on a semi-infinite domain Y ∈ (0, ∞), in the combined limit κ, δ → 0.
Numerical method
We solve the model (2.20) numerically using the method of lines.Specifically, we first transform the model onto two separate fixed domains, setting η = Y/H(T) for the gas-vapour problem, which then holds in η ∈ (0, 1), and setting ξ = (Y − H(T))/(1 − H(T)) for the liquid-dirt problem, which then also holds in ξ ∈ (0, 1).We discretise spatially on these transformed domains, with a uniform mesh, using central differences for diffusive terms and first-order upwinding for advective terms, so that the scheme is overall first order.(The advection for the vapour problem (2.20a), including the artificial advection terms due to the change of variables, is negative; the purely artificial advection in the liquid-dirt problem (2.20b)-(2.20c) is also negative.Upwinding these terms therefore requires forward differences in both cases.)We then use the inbuilt ODE solver ode15s in Matlab for the time stepping.We note that the model is stiff in certain parameter regimes of interest (δ 1 and/or σ 1), and that ode15s is specifically designed for stiff systems.Ode15s is a multistep solver, using numerical differentiation formulas of order 1-5 (Shampine & Reichelt 1997).We make use of our early time asymptotic solution of § 4.1 to initialise our numerical simulations.In particular, the spatial mesh must be sufficiently fine to resolve the boundary layer in the suspended dirt problem at Y = H at early times.Our analysis in § 4.1 suggests we require the number of spatial mesh points N to scale like More efficient solvers might take further advantage of the asymptotic structure of the system and distribute mesh points unevenly through the domain in order to ensure good resolution of the boundary layer while maintaining computational efficiency.However, by making use of our early time asymptotic solution we do not require simulations at particularly small T, and our uniform-mesh formulation suffices.
The slow deposition limit κ 1 and dry clogging
Having stated the model and our numerical solution method, we are now in a position to explore solutions of the model.In this section we focus on the limit κ 1, for which the dirt-deposition time scale is much longer than the evaporation time scale.We expect that the accumulation of suspended dirt due to evaporation and the effects of suspended dirt diffusion to be dominant.
We first consider the leading-order behaviour, taking κ = 0, and show that the evaporation becomes infinitely slow as suspended dirt accumulates.We then allow κ to be small but non-zero, and explore our first clogging scenario, which we term 'dry clogging'.
5.1.
Infinitely slow evaporation when κ = 0 Taking κ = 0, we see from (2.20c) that we have R = 0 everywhere, and thus, D = D 0 , C = C 0 and φ = φ 0 are all constant, equal to their values at R = 0. Taking the δ 1 limit as in § 3, the motion of the evaporation front is therefore governed by (3.4), i.e. (5.2d) To investigate how the accumulation of suspended dirt affects the evaporation rate, we consider the additional limit of σ 1 so that the diffusion of suspended dirt is fast.In this case, we see from (5.2) that θ(T) is uniform, and so, for overall conservation of dirt, we must have .
(5.3) Substituting (5.3) into (5.1)we obtain the single equation for H(T), (5.4) As discussed previously, the evaporation shuts down when θ = θ so that f ( θ) = α, since then H T = 0.At this point we see from (5.3) that H = 1 − θ IC / θ.Numerical solutions of the model (2.20) (with κ = 0, δ = 10 −3 ) are shown in figures 4(a) and 4(b), and compared with the solution of (5.4) for the limit of σ → 0, with good agreement for σ = 0.1 and smaller.We take the functional form f (θ ) = 1 − θ for these simulations, and fix α = 0 (so that θ = 1).In figure 4(c) we show solutions of (5.4) for various θ IC .We see that, for larger θ IC , the evaporation is slower, with the evaporating interface H moving more slowly into the domain.In particular, we note that when θ IC = 0, the evaporation is completed (with H = 1) in finite time ≈ 0.36, (5.5) (from (5.4) with θ IC = 0), whereas for θ IC > 0, we see that H appears to take infinite time to reach 1 − θ IC / θ .We investigate this late time behaviour (within the σ 1 limit) by considering the expansion where c 1 is small and H = O(1) is positive.Assuming that f is continuous at θ = θ, on substitution of (5.6) into (5.4)we find that (retaining only leading-order terms on either side) (5.7) So long as the gradient of the function f is bounded at θ, we may Taylor expand the right-hand side of (5.7) and, thus, find that, to leading order in c, since f ( θ) = α by definition.Thus, H decays exponentially to zero as T → ∞ (since f ( θ) < 0) and the evaporation takes infinite time.(Similarly, if f ( θ) = 0 but higher derivatives are non-zero then H has polynomial, and still infinite-time, decay.)We might ask if there are sensible choices for f (θ ) that result in finite-time completion of the evaporation.In order for the evaporation to complete in finite time, we see that the gradient of f must be unbounded at θ .For instance, if f ( θ(1 − cH)) = α + A(cH) 1/n , for n > 1 and some constant A, then by substituting this expansion for f into (5.7) and integrating the resulting ODE for H in time T, we find that , (5.9) and so H reaches zero in finite time.However, this finite-time drying requires that the gradient of f is unbounded at θ , which is physically unreasonable, not least because θ (defined as the value of θ for which f = α) depends on the atmospheric vapour density ρ a via the value of α.For physically reasonable functional forms f (θ ), we therefore expect a bounded derivative at θ, and thus, that the evaporation becomes unboundedly slow as θ → θ.
This analysis is for the case σ 1.For larger σ , we see from figure 4(a) that the evaporation rate is slower.As we see in figure 4(b), this is because suspended dirt accumulates near the evaporating interface rather than quickly diffusing through the domain, and with higher values of θ| H , we see from (5.1) that the evaporation rate is reduced.Numerically, we see that, for non-negligible σ , we still have H → 1 − θ IC / θ in infinite time.Indeed, although for larger σ the evaporation is additionally slowed by the diffusion of the suspended dirt away from the evaporating interface, at late times when the evaporation becomes infinitely slow, the diffusion of dirt does not limit the drying process.
In summary, for κ = 0, the drying takes infinite time to complete and, as drying occurs, the suspended dirt concentrates in a layer at y = 1.The drying never fully stops (although it becomes infinitely slow).By contrast, we see in § 5.2 that if κ 1 is non-zero then the dirt begins to deposit at late time, and this causes the drying to completely stop in finite time (which we refer to as clogging).
Dry-clogging behaviour for small but non-zero κ
The analysis in § 5.1 assumed that κ = 0 so that there was no deposition of dirt at all, and we saw that, in this case, there is an infinitely long drying time as θ → θ.However, in reality we might instead have κ 1 small but non-zero.In this case, we would expect the same behaviour as in § 5 initially (over an O(1) time), but then during the slow evolution towards θ → θ, the deposition will begin to become important.Dirt is slowly deposited, and the deposited dirt layer, thickness R, grows.From the microscale geometry, we see that, when R = R clog := 0.5 − r 0 , the dirt layers on neighbouring solid circles meet, and the pore-scale liquid region ceases to be connected.This means that the effective diffusivity D(R clog ) = 0.In particular, if R = R clog then vapour cannot be transported through the porous material, and thus, evaporation ceases.We define 'clogging' to be this situation when R = R clog at Y = H(T) at finite time T, and thus, the evaporation is stopped.
To investigate this, we look at the behaviour of the system on the long time T = T/κ over which R varies.With this change of variables in (5.1), we see that H satisfies (5.11) (5.12) along with the boundary conditions (2.20f ) and (2.20h), ) At leading order in κ, we see from (5.12)-(5.14) that θ = θ is uniform, and thus, in Y > H, from (5.11) we find that R = ( θ − β) T (5.15) is spatially uniform.Clearly R = R clog after time T = R clog /( θ − β), or in the original time variable, at + O(1). (5.16) We note that, during this late, O(1/κ) time, the position of the evaporating interface, H, and the O(κ) deviation of θ from θ may be found by going to next order in κ in (5.10) and (5.12), and matching to the early time behaviour in T (or O(1) time behaviour in T).Thus, R reaches the clogging point R clog in finite time, and thus, the system clogs.We term this type of clogging 'dry' clogging because, to leading order, θ = θ everywhere ahead of the clogging front.Since we expect θ ≈ 1 (indeed we take θ = 1 in our numerical simulations), there is therefore a negligible amount of liquid left trapped in the porous material by the clogging.This dry-clogging behaviour is illustrated in figure 5(a).Results of a numerical simulation of (2.20) for κ = 0.04 are shown in figure 6.As expected, we see in figure 6(b) that H varies from zero to around 1 − θ IC / θ over an O(1) time, while R at the interface Y = H (where R is maximised in space at that time T) remains closer to zero during the time for which H varies. Subsequently, there is a longer O(1/κ) time over which H remains nearly stationary, since θ ≈ θ everywhere in Y > H, while R (at Y = H) increases linearly to R clog = 0.3.The prediction (5.16) gives T end = 7.5 for the parameter values used in figure 6(b), which is seen to be fairly accurate, although a slight underestimate, as this does not take into account the early time (in T, or O(1) in T) stage.
The dry clogging that we have described in this section always occurs for κ 1, but it may also occur for κ = O(1), when the deposition rate is of the same order as the evaporation rate.Indeed, the model (2.20) must always dry clog for κ > 0, even if β = 0.This is because, if β = 0, θ can never reach zero even for large κ, and can only decay exponentially towards it.However, if κ 1 is sufficiently large that θ is very close to zero by the end stages of the drying, the dry clogging occurs at a negligible distance from the end of the domain, H = 1, and is not physically meaningful.We discuss this more in § 7 below.Furthermore, if β > 0 then we expect θ ≥ β for all time (so long as this is true initially).In this case we would certainly anticipate much more prominent dry-clogging behaviour, for a wider range of parameters, although to investigate this thoroughly is beyond the scope of the present study.
The fast deposition limit (κ 1) and wet clogging
We now consider the limit of κ 1, in which the deposition of dirt occurs much faster than the motion of the evaporation front.Since the deposited dirt layer grows quickly, it can become large enough to significantly affect the effective diffusivity D and porosity φ, if θ IC is sufficiently large, impacting on the drying dynamics.In particular, the deposited dirt-layer thickness R may reach the maximum radius R clog = 1/2 − r 0 , at which the diffusivity D = 0, and the system is clogged early in the domain, when θ is not θ in Y > H(T), so that a non-negligible quantity of liquid is trapped by the clogging.We refer to this type of clogging as 'wet clogging' as, compared with the dry-clogging behaviour discussed in § 5.2, a non-negligible amount of liquid is trapped by the clogging.This wet-clogging mechanism is illustrated in figure 5(b).
Large-κ behaviour
To understand the deposition (and potential clogging) behaviour when κ 1, we change to the fast time scale over which R varies, by setting where t = O(1).At such early times the evaporating interface is close to the surface of the porous material, and we rescale (6.2a,b) for Y < H(T) (assuming that σ κ 1), so that (3.3) becomes To leading order in (κσ ) −1 1, we have a plane-autonomous deposition system: We refer to the system (6.6) as the outer problem, which holds away from the evaporation front in the majority of the domain.(At the evaporation front there must be a boundary layer, which we discuss later.)Specifically, the initial conditions θ = θ IC > β and R = 0 at t = 0, imply that both θ and R are independent of Y for all t, and, recalling that which is independent of time t.Equation (6.7) may be interpreted as an expression of overall conservation of dirt: since there is no transport of dirt on this time scale, the total suspended dirt in the liquid and deposited dirt in the layer, must remain constant in time.
We could use (6.7) in (6.6) to find implicit expressions for the spatially uniform R and θ, although this is not particularly illustrative.Instead, we use (6.7) to plot the phase plane in the outer region in figure 7. The system begins at R = 0, θ = θ IC .If θ IC > β, we see from (6.6b) that R increases in time and, from (6.7), that θ decreases towards θ = β.If instead θ IC ≤ β, then R remains zero and θ remains at its initial value (as any deposited dirt immediately re-suspends, from (6.6b)).We note that the system clogs if R reaches R clog before θ reaches β.We see that this occurs for sufficiently large initial suspended dirt concentrations, namely (from (6.7)) if Although this analysis shows that the system certainly clogs for θ IC > θ crit IC (in this limit κσ 1), we expect that the system will in fact clog for lower values of θ IC too, since we expect there will be higher θ and, therefore, faster deposition near to the evaporation front at h.Indeed, in a boundary layer of width O(1/ √ σ κ), dirt accumulates due to the motion of the evaporation front.By making the change of variables we find that, in the boundary layer z ∈ (0, ∞), (6.10b) while at the evaporation interface z = 0, where R out and θ out are the values in the outer region, as described above.This boundary layer system (6.10),coupled with (6.3), describes the motion of the evaporation front, accumulation of suspended dirt ahead of it and the deposition of dirt.We note that dirt accumulation, transport and deposition all balance in the boundary layer.
We show a solution with κ = 100 in figure 8. Figures 8(a)-8(d) show the spatial profiles for ρ, θ and R at four successive times, while the motion of the evaporation front H(T) is shown in figure 8(e), with the time points of the plots (a-d) marked as red circles.At early times, in figures 8(a) and 8(b), we clearly see the boundary layer structure in the θ and R plots in Y > H, with both R and θ uniform in the rest of the domain.As time progresses, we see that R increases while θ decreases, as suspended dirt becomes deposited onto the solid structure.By the time shown in figure 8(c), we see that θ is close to zero everywhere: the deposition has nearly finished.Since κ 1, this occurs when the evaporation front is still only a short O( √ κ) distance into the domain.After this time, θ(≈0) and R are both constant in Y > H, and the evaporation front travels to the bottom of the domain, resulting in a fully dry porous material with non-uniformly deposited dirt.Indeed, we note that the combination of dirt accumulation, diffusion and deposition in the boundary layer results in an internal spatial peak in the final thickness, R, of the deposited dirt layer (for Y < H) near the top of the domain.Since R is higher here, the effective diffusivity of the vapour is correspondingly reduced.This can be seen in the non-monotone gradient of ρ in figure 8(d), where the ρ profile is steepest at the peak value of R. The reduced diffusivity here limits the drying rate for the remainder of the process.
The fact that we obtain this internal peak in R at early time due to the boundary layer accumulation, diffusion and deposition of dirt means that our estimate for the clogging criterion (6.8), which assumes that dirt deposits in a spatially uniform way, must be an upper bound on the true critical θ IC : we expect to have clogging at θ IC lower than the critical value given by (6.8).Indeed, in figure 9 we show a simulation for the same parameter values as in figure 8, except that we take a larger initial suspended dirt volume fraction θ IC = 0.6, which is still lower than the estimate of θ crit IC = 0.755 given by (6.8), for the chosen parameter values.Nevertheless, we see that the system indeed clogs early in the domain.Both liquid and suspended dirt are trapped ahead of the clogged point, since θ < 1 in Y > H.The clogging happens at an O(1/ √ κ) distance into the domain, at H ≈ 0.07, and after an O(1/κ) time, as predicted by our analysis above.
We note that the motion of the evaporation front shown in figure 9(b) no longer follows a √ T behaviour.Instead, we see that the speed H T of the evaporation front is not smoothly decreasing.Evaporation is fast to start with as R is small and the vapour has only a short way to travel to the surface of the porous material.Then the evaporation front begins to slow down, as both θ| H increases due to accumulation and the effective diffusivity D decreases, as R begins to increase.As the clogging point is approached, H T increases again, because the porosity φ is decreasing, so that there is less liquid in the pore space to be evaporated since so much of the pore space is occupied by the deposited and suspended dirt.Similar time-varying behaviour of H T is visible at early times in the case shown in figure 8(e) (inset), when the system did not clog.
Evidently, wet clogging can occur at initial suspended dirt volume fractions θ IC significantly below the estimate for the critical value given by (6.8).To better understand the clogging criterion, we must better understand the behaviour in the boundary layer at Y = H.However, the nonlinearity of the boundary layer equations (6.10) makes them . Numerical solution of (2.20) for large deposition rate κ = 100.We also take f Wet clogging: numerical solution of (2.20) for large deposition rate κ = 100 and θ IC = 0.6.We also take f (θ) = 1 − θ, α = 0, β = 0, σ = 1, ν = 0.5, δ = 10 −3 and r 0 = 0.2.The green dashed line shows the clogging point, R clog = 0.5 − r 0 , which is attained at T ≈ 5.1 × 10 −3 in this simulation.The profiles in (a) are at the time when the system clogs with intractable to further analytical progress.Instead, we study a paradigm problem in the next section, which is a simplified version of the large-κ problem, but which still captures the essence of the wet-clogging behaviour.
The clogging criterion for a paradigm problem
In § 6.1 we saw numerically that, for large κ, the deposited dirt-layer thickness R is non-uniform near Y = 0, increasing from zero to R max := max T (R| H ) that is larger than the final value of R attained in the outer region.This means that wet clogging occurs for lower values of θ IC than predicted by the outer region estimate (6.8).Since the spatially uniform deposition dynamics in the outer region do not capture this non-uniformity in R near to Y = 0, we consider the behaviour in the boundary layer at Y = 0, where there is a balance between all of the accumulation, diffusion and deposition of dirt processes, in order to derive an improved criterion for wet clogging.
However, the boundary layer equations (6.10) are intractable analytically, due to the coupling between variables and the nonlinear terms.In order to build intuition for what determines the size of the peak R max , in this section we investigate a paradigm problem, with a different functional form for f (θ ), and unphysical simplifications of D and the bulk deposition term.With these (non-systematic) simplifications, we solve the boundary layer equations (6.10) explicitly, and hence, determine R max analytically.In this way, we determine a criterion for wet clogging (given by R max ≥ R clog ), and build intuition for the more general case.
We assume, for simplicity, that β = 0 and α = 0. Additionally, we suppose that r 0 is close to 1/2, so that R r 0 (since the circular inclusions are close together, and only thin deposited dirt layers are possible before clogging occurs).Then φ ≈ φ 0 and C ≈ C 0 are approximately constant, even for R approaching R clog .We make the additional approximations Finally, we alter the deposition term in (6.10) by replacing the factor θ * − θ with the constant value θ * .With these choices of functional forms (which we emphasise are not a true limit of the full problem), the paradigm boundary layer problem is, while θ < 1, while at the evaporation interface z = 0, where the outer solution R out , θ out satisfy Solving (6.12), we see from (6.12a) that the evaporation front is simply (6.13) where, for simplicity of notation, we define Furthermore, the outer region phase plane equations (6.12f )-(6.12g)decouple, with the explicit solution (6.15a,b) where, again for simplicity of notation, we define The boundary layer equations (6.12b)-(6.12e)are therefore (6.17a) while, at the evaporation interface z = 0, Thus, under all of our simplifying assumptions, the equations for θ and R decouple: we may solve the linear system (6.17a)(left), (6.17b) and (6.17c) (left) for θ, before then computing R. Specifically, we look for a solution for θ of the form is a similarity variable.Thus, in the boundary layer the far-field deposition solution θ out = θ IC e −Ct is modified by an 'accumulation factor' F, which describes how the evaporation causes suspended dirt to accumulate near the evaporating interface.Substituting this form of θ into (6.17a)(left equation), (6.17b) and (6.17c) (left equation), and solving the resulting ODE system for F(η), we find that The factor 1 − √ πBe B 2 erfc(B) is always positive (although it approaches zero as B, or equivalently σ , approaches ∞), and so the accumulation factor F > 1 (and F → 1 as η (or z)→ ∞).Thus, as we would expect, the value of θ is higher in the boundary layer than in the far field.The large θ in the boundary layer results in greater deposition occurring there, and so R increases compared with the solution as z → ∞.The problem (6.17a) (right equation) for R is first-order hyberbolic, and -given the form (6.19) -may be solved using the method of characteristics.Specifically, we find the characteristic curves take the form z = z 0 − 2B √ Dt, where z 0 parameterises the initial data R = 0 at t = 0. (The shape of the characteristic curves mean we do not require data for R at z = 0.) The solution R, in terms of z and t, is given by The solutions (6.19) and (6.20) of the paradigm model (6.17 evaporation front.Evaluating (6.20) at z = 0, we have In order to find the maximum value R attains, we simply maximise R| h over t.We find that there is a single maximum for positive t, at the critical time t * = τ 2 * /C, where τ * = τ * (B) The solution τ * of (6.22) is shown as a function of B in figure 11(a).We see O(1) variation of τ * when B varies over six orders of magnitude.Indeed, τ * → √ 3/2 as B → ∞ (or σ → ∞, say, the limit of slow dirt diffusion), whilst τ * grows very slowly as B → 0, behaving like the growing root of τ * e −τ 2 * = √ πB 2 for B 1, namely (6.23)where W −1 is the lower branch of the real Lambert W function.
From the critical time t * , we can compute the value that R attains at its maximum, namely where We note from (6.14a,b) that B = O( √ σ ), and so R max and G increase linearly with σ as σ → ∞.Thus, we see that, no matter how small the initial suspended dirt level θ IC , for sufficiently slow dirt diffusion (sufficiently large σ ), we will find that R max exceeds R clog and the system will clog.Indeed, this expression (6.24) for R max gives the clogging condition for this paradigm setting: the system clogs when R max ≥ R clog .Using (6.16), this clogging criterion is We show this critical initial suspended dirt volume fraction (6.27) as a function of the suspended dirt diffusion rate σ in figure 12, along with the small-and large-σ limits (which follow directly from the small-and large-B behaviour of G discussed above), namely We note that, since G ≥ 1 and R clog > 0, this θ crit IC , given by (6.27), for the paradigm problem is strictly smaller than the upper bound (6.8) (that essentially assumes a uniform deposited dirt layer), since, in the case β = 0 that we have assumed for the paradigm case, the upper bound (6.8) may be rearranged to be written in terms of φ 0 = 1 − πr 2 0 , C 0 = 2πr 0 and R clog = 1/2 − r 0 , becoming Thus, as expected, the non-uniformity of the dirt deposit in the boundary layer at the evaporation front results in clogging for lower initial suspended dirt volume fractions than if the dirt were to deposit uniformly.We note that our expression (6.27) for θ crit IC in this paradigm case is independent of κ, since we have taken the leading-order behaviour as κ → ∞.Since we require the boundary layer structure, our paradigm for the clogging criterion is only valid when the width 1/ √ σ κ of the boundary layer is small, and so only for σ sufficiently large that σ 1/κ (although since we assume κ 1 this is not particularly restrictive).Additionally, we see that the expression (6.27) is independent of D, and hence, of D 0 .From this, we learn that it is the relative diffusivity of the suspended dirt and the liquid vapour, captured through σ that is important, and not how both are equally affected by the presence of the porous material (recall that D 0 is the effective diffusivity).Of course, the depth y into the porous material at which the clogging occurs does depend on D. Specifically, the clogging depth when θ IC = θ crit IC takes its critical value is Curiously, although this position given by (6.30) depends on κ and D 0 , it is only weakly dependent on σ , since τ * (the solution of (6.22)) was seen to have very weak dependence on B ∝ √ σ .We considered a simplified, paradigm version of the problem in this section, in order to make analytic progress.Although this analysis does not directly relate to the full drying model, we may extrapolate some general conclusions.Firstly, the qualitative wet-clogging behaviour seen in the full model, characterised by an internal peak in R, does not rely on the variation of D, φ and the evaporation-front speed h T with R and θ (in the paradigm case we supposed all these were constant).Instead, the important mechanism, captured by the paradigm model, is the variation of the rate of dirt deposition with θ, along with the fact that θ is spatially non-uniform in the boundary layer at the evaporation front, determined by a balance of all three mechanisms of diffusion, accumulation due to liquid evaporation and the deposition.
In the following section we generate bifurcation diagrams showing parameter regimes for which the system clogs.We compare these numerical results with the predictions of the paradigm model as appropriate.
Clogging parameter regimes
Having built understanding of the two mechanisms by which the porous material may clog, in this section we compute bifurcation diagrams numerically in order to quantify the parameter regimes for which clogging occurs.
Firstly, we consider the case κ 1 for which the dirt-deposition rate is high relative to the evaporation rate.As in § 6, we do not expect the system to dry clog, but instead to exhibit wet clogging for sufficiently large θ IC and σ .In figures 13(a) and 13(b) we show the numerically computed regions of parameter space for which dry clogging occurs, for two different values of the microscale-geometry parameter r 0 .For both, we observe that there is a well-defined critical suspended dirt volume fraction θ crit IC above which the system clogs, and below which there is no clogging and the evaporation front reaches H = 1.We see that θ crit IC is a monotone decreasing function of σ .The estimate (6.8) is indeed seen to be an upper bound for the numerically computed θ crit IC , and is most accurate for small σ , when the diffusion of dirt is fast and so dirt deposition is approximately uniform.For r 0 = 0.45 (figure 13a), where R clog = 0.05 is fairly small and φ and C do not vary much with R, we see that the prediction (6.27) of the paradigm model is, in fact, a reasonable approximation for the full system, despite the fact that the paradigm model is not a real limit of the full model.In particular, for large σ , we observe wet clogging for very small initial suspended dirt volume fractions.For r 0 = 0.2 (figure 13b), the paradigm model prediction of θ crit IC is a poor approximation of the full system.In figure 14 we investigate the effect of the deposition rate κ and the initial condition θ IC on the clogging behaviour.We simulate the model (2.20) for each set of parameter values (θ IC , κ), and in figure 14(a) the colour indicates the position H end of the evaporating interface at the time when the simulation terminated (so H end = 1 if there was no clogging and the evaporation completed, while H end < 1 if the system clogged).In figure 14 We note that the κ axis is on a log scale in figure 14.For large κ, we see that there is a critical θ IC ≈ 0.65, which appears to be largely independent of κ, above which we have wet clogging and below which the system does not clog.A non-negligible amount of liquid remains trapped in the pore space when the evaporation terminates.This wet-clogging behaviour is as discussed in § 6.The upper bound on the critical initial suspended dirt volume fraction θ crit IC , given by (6.8) and shown by the red dashed curve, is seen to be a significant overestimate, even for the relatively small value σ = 0.1 (as in figure 13, we expect (6.8) to be most accurate for small σ ).We see that, as κ increases, the wet clogging occurs earlier (H end is smaller) and correspondingly more liquid remains trapped in the pore space.
For small κ, we observe dry-clogging behaviour, as discussed in § 5, with H end < 1 but with a negligible amount of liquid trapped in the pore space, since θ ≈ 1 for Y > H in this case.For κ 1, we see that dry clogging occurs for all θ IC > 0, to varying degrees, with H end close to one for small θ IC , but actually very small for larger values of θ IC .The transition from the no-clogging/wet-clogging regime for κ 1 to the dry clogging for κ 1 is gradual.Indeed, for κ = O(1), for which we were unable to make analytic progress, it is not clear how to classify the system.For small θ IC , there is some dry clogging, with H end ≈ 1 and the remaining liquid volume very small.Moreover, for larger θ IC , the behaviour could be considered to be either dry or wet clogging, with H end around halfway through the domain, and a fairly small amount of liquid remaining trapped in the pore space.
Discussion and conclusions
We have derived and analysed a model for the drying of liquid from within a porous material and the associated deposition of impurities within the pore structure.By beginning with a pore-scale model and systematically homogenising, we incorporated delicate couplings between the dependent variables, including the effect of a growing layer of deposited dirt on the porosity, the diffusivity of both dirt and vapour, and on the deposition rate of the dirt.We explored the relevant limit where the vapour density is much smaller than the liquid density (δ 1); in this case, the vapour-transport problem was reduced to a single equation for the motion of the evaporation front.Our resulting equation is valid in the physically relevant limit where the vapour transport through the porous material limits the evaporation.This is different to prescribing an evaporation rate, which is a common approach in the literature, and which is only valid when the vaporisation of the liquid molecules is the limiting mechanism.
The accumulation of suspended dirt at the evaporating interface during drying was shown to reduce the evaporation rate, since we imposed a dirt-dependent saturation vapour density at the evaporating interface.We also saw that, in the limit of slow suspended dirt diffusion, the transport of the dirt away from the evaporating interface limits the evaporation rate.The thickness, R, of the deposited dirt layer was seen to vary spatially within the porous material.For slow deposition rates κ, R increased monotonically into the porous material, with the majority of the dirt concentrated at the end of the porous material.Conversely, for large κ, we observed an internal peak in R a short O(1/ √ κ) distance from the external surface of the material, and a uniform-thickness deposit through the majority of the remaining material.These spatial non-uniformities in R were shown to result in two distinct clogging mechanisms, in distinct regions of parameter space.The first clogging mechanism was dry clogging, where deposition is slow, and suspended dirt is pushed further and further into the material as the evaporation front passes through the domain, until there is insufficient space for it all to deposit and the system clogs.
A negligibly small amount of liquid is trapped in the system during dry clogging.By contrast, we found that wet clogging, defined as clogging when both liquid and suspended dirt are trapped in the porous material, occurs at sufficiently high dirt-deposition rates κ and sufficiently slow suspended dirt diffusion rates σ −1 , such that the internal peak in R is too high, and the deposited dirt layer clogs the pore space.We constructed a simplified paradigm model in the large-κ situation, which captured the key mechanisms of coupled accumulation, diffusion and deposition of dirt in a boundary layer at the evaporating interface, and derived a wet-clogging criterion.
For industrial drying scenarios, it may be important to control the dirt-deposition profile.In particular, it may be important to obtain as uniform a deposited layer through the material as possible, for instance, if it is a dye or ink pigment that is being deposited.In the drying of filters and textiles after cleaning, clogging of the system should be avoided as much as possible, since a clogged filter can no longer perform its function.The drying rate might more easily be controlled than the diffusivity or deposition rate of the dirt (perhaps by controlling the ambient temperature or humidity) in order to avoid clogging-prone parameter regimes.Specifically, so long as the total initial amount of dirt is sufficiently low that wet clogging will not occur, the drying rate should be kept slow (i.e.κ should be made large), in order to avoid dry clogging.
For our numerical simulations, we chose a simple linear dependence of the saturation vapour density on the suspended dirt volume fraction f (θ ) = 1 − θ, but this should be investigated further, since it is a key mechanism by which the accumulation of suspended dirt affects the evaporation rate.We also focused the majority of our analysis in the case of α = 0 so that the ambient humidity is very low, and the case β = 0, so that the dirt cannot re-suspend into the liquid once deposited.The effect of non-zero α and β should be further investigated.In particular, we would expect dry clogging to be more prominent in the case β > 0, even for high deposition rates κ, since the suspended dirt volume fraction would not decay to zero ahead of the evaporation front in this case.We also used a two-dimensional microstructure, with circular solid inclusions.Three-dimensional microstructures should be investigated, such as an array of spheres, which would result in different functional forms for C, φ and D. In particular, the liquid region would remain connected when the dirt layers growing on neighbouring spheres met, although continuing growth of the dirt would eventually result in clogging in a similar way to our two-dimensional case.We expect qualitatively similar behaviour for alternative micro-geometries to our two-dimensional circles, including the possibility of both dry and wet clogging in the appropriate parameter regimes.The model and homogenisation analysis could be extended to other geometries, such as hexagonally packed cylinders or square solid inclusions, as well as more general pore-scale geometries by using a level-set description of the microscale dirt-liquid interface, as in, for instance, Richardson & Chapman (2011).We should additionally investigate our model in higher macroscale dimensions, so that the evaporating interface is at Y = H(X, T).In particular, in the slow-dirt-diffusion case σ 1, which is analogous to a Stefan problem with constitutional supercooling, it is possible that the evaporating interface may become unstable.
A key assumption of our modelling was that the liquid remained stationary and did not flow.This resulted in a sharp evaporating interface separating the liquid/dirt and gas/vapour occupied regions of the porous material.In reality, a capillary-driven flow of liquid towards the surface of the porous material could draw suspended dirt to the surface as well, and we might anticipate even higher peaks in the deposited dirt-layer thickness at or near the surface, increasing the likelihood of wet clogging.Incorporating such capillary flows will be an important area for our future work.
The drying model derived in this paper captures many subtle couplings between the evaporation, accumulation, transport and deposition of dirt, and the transport of vapour out of the porous material during drying.The model itself and our subsequent analysis constitutes an important step towards an accurate prediction of deposited dirt profiles and clogging behaviours, of particular relevance in filtration and the textile industry.
Figure 1 .
Figure1.Schematic showing the evaporation front at y = h(x, t) moving through the pore space (length scale d) of a porous material (of depth l d), with dirt depositing in a layer of thickness R(x, y, t) on the circular solid inclusions with radius r 0 .The unit normal to the solid or dirt boundary of the pore space is n s , while the evaporating interface has unit normal m.
Figure 10 .
Figure10.The black curves are the analytic solution (a) θ given by (6.19) and (b) R for Y > H given by (6.20), of the paradigm problem (6.17), at various times through the drying, with arrows showing increasing time.The red curve in (b) shows the final deposited dirt layer after the drying is completed.We have taken parameter values κ = 100, ν = 0.5, σ = 1, θ IC = 0.1 and r 0 = 0.4.
Figure 11 .
Figure 11.The maximum of R for the paradigm problem.(a) The solution τ * of (6.22) as a function of B, with the large B limit √ 3/2 (red dashed) and the small-B limit, namely the (non-negligible) root of τ * e −τ 2 * = √ πB 2 (green dashed).(b) The (scaled) maximum value of R max , given by (6.25), against B. For small B, we have G ∼ 1 while, for large B, we have G ∼ e −3/2 ( √ 6 − 1)B 2 (red dashed).
Figure 14 .
Figure 14.Clogging behaviour of (2.20) as κ is varied.(a) The end position H end when the evaporation terminates, so H end = 1 if the drying is complete without clogging, H end < 1 indicates clogging.(b) The total volume of liquid remaining in the pore space when the evaporation terminates.In both (a,b), the red dashed line indicates the upper estimate (6.8) for the wet-clogging criterion.Throughout the figure we take σ = 0.1, f = 1 − θ , α = 0, β = 0, ν = 0.5, r 0 = 0.2 and δ = 10 −3 .
(b), for the same set of model simulations, the colour indicates the volume of liquid remaining when the simulation terminates, namely liquid volume remaining = 1 Y=H end φ(1 − θ) dY.(7.1) | 18,475.8 | 2024-05-10T00:00:00.000 | [
"Environmental Science",
"Mathematics",
"Materials Science"
] |
PHYSICAL AND MECHANICAL PROPERTIES OF HEAT TREATED WOOD FROM Aspidosperma populifolium , Dipteryx odorata AND Mimosa scabrella
Heat treatment improves some wood properties namely: equilibrium moisture, dimensional stability and durability and mechanical properties. In this study, the heat treatment was applied to woods of three natural species from Brazil: Aspidosperma populifolium (peroba mica), Dipteryx odorata (cumaru) and Mimosa scabrella (bracatinga). The woods were heated in an oven under vacuum and under nitrogen, at 180, 200, and 220°C for one hour. The untreated and heat-treated woods were characterized in relation to equilibrium moisture content, basic density, shrinkage, Janka hardness, and bending MOR and MOE according to NBR 7190 standards. All the thermal rectified woods showed a reduction in the hygroscopic equilibrium content, especially when the heating was under vacuum from 13-15% in the untreated woods to 1-3% for vacuum treatment at 220 °C. The dimensional stability was improved to only a small extent e.g . volumetric shrinkage tended to decrease with increasing temperature. The mechanical properties were affected differently for the three wood species. Heat-treated cumarushowed increased Janka hardness, MOR and MOE; and peroba mica increased MOR and MOE but not Janka hardness; while bracatinga was less influenced by the heat treatment.
INTRODUCTION
Wood is a material with excellent mechanical behavior, namely considering its performance-todensity ratio, but also with some undesirable properties such as hygroscopicity, anisotropy, dimensional instability and biodegradability (Tomak et al. 2011, Miller 1999. Thermal modification processes are used to improve the wood quality and are especially targeted to increase durability and dimensional stability, and decrease equilibrium moisture content (Esteves and Pereira 2009). The heat-treated wood can therefore be applied for higher value end-uses such as surfacing, floors, windows and doors, musical instruments, boats, and general outdoor uses (Gunduz 2009).
Thermal modification is performed by heating the wood at temperatures between 180 and 260 °C, usually in the absence of oxygen or with air deficiency (Homan 2004). In these conditions, there are chemical changes e.g. degradation of hemicelluloses and extractives, to an extent that depends on the wood species and on the treatment intensity, for example under 140 °C only slight changes occur while above 260 °C the substantial degradation of the structural components lowers the wood properties Thermal modification has been applied to different species, mostly hardwoods and softwoods from temperate regions. Few species from tropical and subtropical regions have undergone thermal treatment: Pinus caribaea var. hondurensis, between 120 and 180°C (Borges and Quirino 2005), Eucalyptus grandis between 120 and 200 °C (Brito et al. 2006) and Corymbia citriodora between 200 and 220 °C (Nunes 2009), while Eucalyptus grandis, Pinus taeda and Tectona grandis were heat treated at 160 °C at industrial scale (Lengowski 2011). The Brazilian native species Simarouba amara, Sextonia rubra and Cariniana micrantha were also treated at 150 and 200 °C to reduce surface color variation (Gouveia 2008).
This work studies three species occurring naturally in Brazil and with economic importance for the furniture and flooring industry.
Mimosa scabrella Bentham (common name bracatinga) is a native forest species in Brazil, occurring in cold and wet areas, mainly in the states of Minas Gerais, Paraná, Rio de Janeiro, Santa Catarina, Rio Grande do Sul and São Paulo. The wood has a medium density (510 to 610 kg m -3 ), low durability when in contact with the ground, and low permeability towards preservative products (Sturion andSilva 1989, Inoue et al. 1984). It is used for firewood, charcoal, plywood, particleboard and lightweight containers and, more recently, started to be introduced in furniture and flooring (Baggio et al. 1986, Baggio andCarpanezzi 1998).
Aspidosperma populifolium (common name peroba mica) is distributed naturally in the Amazon region and center-west of Brazil, in the states of Amazonas, Pará, Rondônia and Mato Grosso. The wood is density (730 kg m -3 ) and used in construction, as beams, rafters and slats for floors, carpentry and high quality furniture manufacturing.
Dipteryx odorata (common name cumaru) is frequent in the states of Acre, Amapá, Amazonas, Pará, Rondônia and Mato Grosso, as well as in neighboring countries like Guyana, Venezuela, Colombia, Bolivia, Peru and Suriname. The wood is very dense (950-1000 kg m -3 ), tough, highly durable and resistant to cracking when exposed to sunlight, therefore is suitable for solid flooring, stair treads, furniture, and pool decks (Loureiro et al. 1979).
In this work the wood of these three species (Aspidosperma populifolium, Dipteryx odorata, Mimosa scabrella) was thermally treated and the effect of temperature and treatment methods (vacuum and nitrogen environments) was evaluated in relation to equilibrium moisture, basic density, shrinkage, Janka hardness and bending strength. The aim is to improve the wood properties of these three species in order to target them towards high-quality solid timber products.
Samples
The commercial woods: bracatinga (Mimosa scabrella Bentham), peroba mica (Aspidosperma populifolium A. DC) and cumaru (Dipteryx odorata (Aubl.) Wiilld) were used in this study. The boards were donated by Indusparket (Tietê, São Paulo, Brazil), a national manufacturer of hardwood floors. The boards were air dried to moisture content between 12 and 15 %. Test samples were cut with 60 cm x 7,5 cm x 2 cm (length x width x thickness, corresponding to the axial, tangential and radial directions of the wood).
The experiments were conducted in the laboratories of the Panels and Wood Energy (LAPEM), Pulp and Paper (LCP) and Wood Properties (LPM), of the Federal University of Viçosa, Brazil.
Physical and mechanical properties ..: Araujo et al.
Thermal treatment
The wood samples were thermally treated in a Marconi vacuum oven Model MA-027 (São Paulo, Brazil), fitted with temperature and pressure, or vacuum, control. The oven internal chamber has a cylindrical shape (30 cm diameter and 70 cm length) and a support platform where the wood samples are stacked. The oven allowed the control of temperature, the choice of input gas e.g. air or nitrogen, and the variation of the inside pressure by connection to a vacuum pump.
Preceding each heat treatment, the samples were weighed and measured (length, width and thickness) for later calculation of mass loss and volume variation. The mass loss was determined for each sample in relation to its initial oven-dry mass in accordance to the equation 1: wood mass -treated wood mass Mass loss % = × 100 wood mass (Equation 1) where the wood mass is the oven-dry weight of the specimen without treatment (g), and the treated wood mass corresponds to the oven-dry weight of the specimen after heat treatment (g).
The oven was conditioned to the environment conditions under testing, and heated until the treatment temperature was attained; the samples were then rapidly put inside, and heated during one hour at constant temperature. The temperatures tested were 180, 200 and 220 °C. Two oven conditions were evaluated: i) the air filled oven was evacuated with the coupled vacuum pump to a pressure of 0,06 MPa (here called "vacuum" treatment); ii) the oven was filled with nitrogen and then evacuated with the coupled vacuum pump to a pressure of 0,06 MPa (here called "nitrogen" treatment).
Each experiment used six wood samples for each species. After treatment, the samples were removed and cooled to 25 °C, weighed, measured (width, thickness and length) and kept at ambient temperature and humidity. Posteriorly, the samples were weighed and measured weekly until they reached equilibrium moisture content that was attained after eight weeks.
Wood properties
The following properties were determined in untreated and heat treated wood samples: equilibrium moisture content, basic density, shrinkage, Janka hardness and bending strength. The specimens for each test were prepared according to NBR 7190 (ABNT 1997) standards. Before testing, all the samples were conditioned in a climate chamber at 20 °C and 65 % relative humidity.
Hygroscopicity was evaluated by the equilibrium moisture content after stabilization in a controlled environment at 20±2 °C and 65 % relative humidity.
Wood basic density was determined according to adaptation of Tappi 258 om-02 standard, and calculated by dividing the oven dried mass by the saturated volume, using the standard immersion method for volume determination. The dry mass was determined after oven dry at 103 °C until constant weight.
The test specimens were cut to a prismatic shape with a rectangular cross section of 2 cm x 3cm (radial x tangential) and 5 cm length (axial). Eighteen replicates of untreated and treated specimens were used for each test (three different temperatures, two different methods, three species), totaling 108 samples (6 wood samples x 18 replicates) for each test and by species. The other physical and mechanical tests were made with the same number of specimens.
The dimensional stability of the wood is characterized by the properties of shrinkage and swelling that should be determined in the tangential, radial and axial directions. The shrinkage was determined from a saturated state to an oven dry state, andcalculated as percent variation in relation to the saturated state. The volume variation was determined with the dimensions of the samples in the dry and saturated states as in equation 2: WhereS vol is the total volumetric shrinkage (%); V w is the saturated volume (cm 3 ); and V d the oven dry volume (cm 3 ).
The samples were placed in a saturated environment at 20 °C ± 5 °C, until the dimensional variation stabilized (differences of 0,02 mm between two successive measurements were accepted).
The anisotropy coefficient (AC) was calculated by dividing the tangential shrinkage (TC, %) by the radial shrinkage (RC, %), as follows The Janka wood hardness teste was determined by the ratio between the maximum force (in N) applied for a required penetration into the wood of a sphere with 1 cm 2 diametrical section at a depth equal to its radius. The wood specimens were prismatic with a square of 5 cm in cross-section (radial x tangential) and a length (axial) of 15 cm. For the application of the Janka Hardness test, the specimens were glued together with adhesive (resorsinol) in order to have a square area of 5 cm 2 in the sample cross-section.
Bending was characterized by the flexural modulus of rupture (MOR) and modulus of elasticity (MOE), and the results were reported in MPa. The wood specimens for the bending tests had a prismatic square shape of 2 cm across (radial x tangential) and a length of 30 cm (axial). The mechanical test was performed on a computer controlled universal testing machine CONTENCO (Minas Gerais, Brazil). The specimens were supported on two articulated cleavers, with a span of 24 cm, and the load was applied by means of a cleaver directly in the center of the specimen, tangentially to the growth rings. The load was increased gradually and steadily, allowing the measurement of the deflection at each increment until rupture. Table 1 presents the average values of the equilibrium moisture content in the untreated and in the heat-treated samples. The untreated woods had equilibrium moisture values ranging from 13,1% (cumaru) to 15,5% (bracatinga) that decreased significantly upon thermal treatment. In all cases, the heat-treated wood samples significantly differed from the untreated wood (Table 1).
Equilibrium moisture content
The reduction extent of the equilibrium moisture content was influenced by the method used for the thermal treatment: under vacuum, the reduction was much higher than under a nitrogen environment for the same temperature e.g. 3,3% vs. 12,8% for vacuum and nitrogen treatment for bracatinga wood at 180 °C, respectively (Table 1). Temperature also influenced the equilibrium moisture content, although the differences between treatments at 180°C and 220°C were small. The analysis of variance indicated that method and temperature within the method were significant factors of variation of the wood equilibrium moisture content, although in most cases there was no statistical significant variation between temperatures (Table 1).
Physical and mechanical properties ..: Araujo et al. Comparison of the three species showed a similar behavior both regarding temperature as well as the method. It is noteworthy that the three heat-treated woods under vacuum attained very low equilibrium moisture contents: bracatinga 2,7%, peroba mica 1,2% and cumaru 1,3%. In the treatment under vacuum there was already a major effect at 180 °C. Table 1. Hygroscopicity (equilibrium moisture content, %) of bracatinga, peroba mica and cumaru woods after heat treatment at three temperatures (180, 200 and 220 °C) under two methods (vacuum and N 2 ).
Method Wood
Temperature ( Similar results have been observed by researchers studying wood thermal modification; in fact, the decrease of wood hygroscopicity is considered as one of the greatest benefits provided by the thermal wood modification (Hill 2006, Esteves andPereira 2009). The decrease of hygroscopicity in thermally treated woods may be explained by a lower water absorption by the cell walls due to the partial removal of extractives and of hemicelluloses that reduces the hydrophilic hydroxyl groups, with an increase of cellulose crystallinity and of linkages in the lignin matrix (Esteves et al. 2008a, Sundqvist 2004, Boonstra and Tjeerdsma 2006. The treatment under vacuum favored the transfer of volatiles to the gas phase (Candelier et al. 2013), therefore yielding a more hydrophobic wood, as shown by the results in Table 1. The same influence of a vacuum environment leading to a sharp decrease of the equilibrium moisture content with the heat treatment was also reported using the same equipment (Araújo et al. 2014).
Basic density and mass loss
The wood basic density was different between the three species: the lower value was attained by bracatinga (550 kg m -3 ), while peroba mica 610 kg m -3 ; and the higher value was obtained by cumaru, 900 kg m -3 .
The thermal treatment influenced the wood density only in a small extent as can be seen in Figure 1. Bracatinga wood showed a small increase in wood density after the heat-treatment,e.g increased to 58 kg m -3 for the different methods and temperatures; for peroba mica the wood basic density increased slightly to 630 kg m -3 for the heat-treated sample at 220 °C; while in cumaru wood the variation was also small e.g. 890kg m -3 in the sample treated at 220 °C treatment under nitrogen. Temperature is usually considered to induce a decrease of the wood basic density due to mass loss (Boonstra et al. 2007), but response to particular treatment may differ (Tomazello Filho 1985, Garcia 1995. This was found previously for the heat treatmentof eucalypt wood using the same equipment and conditions as in this study (Araújo et al. 2014).
The mass loss was different between the three species and two methods ( Table 2). Mass loss of wood is related to its thermal degradation and appears to be the indicator of the treatment intensity. Since the properties of heat treated wood depend on the mass loss (Weiland et al. 1998) the control of the mass loss during the treatment is a quality factor.
The mass loss values were overall small. Mass loss increased when temperature increased for both methods. This has been shown for instance by Allegretti et al. (2012) who refer mass loss of 1% at160 °C and 6-7% at 220 °C for Norway spruce (Picea abies) and fir (Abies alba). In this work, the values of mass loss found for 220 ° C were lower, presumably due to the heat short treatment time (1 h). In fact, the comparison of data from different authors is often difficult due to differences in treatment duration and other factors involved.
Physical and mechanical properties ..: Araujo et al. The mass loss values were lower for the treatment with nitrogen than for that with vacuum for all temperatures and species. This was an expected result because under nitrogen the medium is deficient in oxygen, there by reducing the oxidative reactions which interfere with the wood components.
Under vacuum, the mass loss remained of small magnitude since all volatile degradation products like acetic acid or furfural were removed as soon as they were formed thereby limiting the degradation of wood polysaccharides and their recondensation through thermoreticulation and crosslinking reactions (Candelier et al. 2013). Table 2. Mass loss (%) of bracatinga, peroba mica and cumaru wood samples after thermal treatment at three temperatures (180, 200 and 220 o C) and using two methods (vacuum and N 2 ).
Wood
Temp ( Table 3 presents the mean values for the axial, tangential, radial and volumetric shrinkage as well as the anisotropy coefficient of the untreated and the thermally treated wood samples. The untreated wood samples showed shrinkage values in accordance with reported values e.g. very small axial dimensional variation and medium shrinkage in the radial and tangential directions. Bracatinga wood is considered to have high shrinkage values which limit its usage (Moreschi 2005, Galvão andJankowski 1985) and the values found here were similar to those reported by Costa et al. (2010) and Stamm (1956). The peroba mica wood showed shrinkage values within the range reported by Logsdon (2008) and Lovatti (2008): 3,6 and 6,2% radial shrinkage; 6,9 and 9,4% tangential shrinkage; 10,4 and 16,6% volumetric shrinkage, with 1,98 and 1,51 anisotropy coefficient. Cumaru wood had shrinkage values in accordance with the reported values of 5,4% radial shrinkage; 8,4% tangential shrinkage, 12% volumetric shrinkage and 1,55 coefficient of anisotropy (IBAMA 1997, IBDF 1988, IBDF 1981, Souza et al. 1997.
Shrinkage of wood
The heat treatment did not change significantly the dimensional stability of the woods, and the differences in shrinkage values with temperatures and treatment methods were not statistically significant.
Physical and mechanical properties ..: Araujo et al. Table 4 shows the values obtained for the Janka hardness perpendicular to the grain for the untreated and heat treated wood samples. The untreated woods showed differences in hardness with cumaru wood as the hardest and bracatinga as the less hard (105,6 MPa and 46,7 MPa respectively).
Janka hardness
The effect of heating on the wood hardness depended on the species (Table 3). The thermal treatments showed no significant influence on the Janka hardness of bracatinga and peroba mica woods. On the contrary, cumaru wood significantly increased its hardness with the heat treatment, regardless of the heating method or the temperature, which did not impact significantly on the results. The increase of strength and hardness of thermally treated wood is attributed to chemical condensation between polysaccharides and lignin (Sundqvist 2004). Different reports on Janka hardness variation in thermally treated woods have been given in the literature. Gunduz et al. (2009) observed that with increasing temperature and duration of heat treatment, Janka hardness decreased; while the Finnish Thermowood Association (FTA 2003) indicated that it increases with temperature, and higher treatment temperatures will influence more the wood hardness (Tuong and Li 2011).
Bending resistance
The heat treatment of the woods influenced their bending behavior, but the sign of the variation depended on the species and treatment type. Bracatinga showed a reduction in the modulus of rupture for the treatment under vacuum but an increase for the treatment under nitrogen, e.g. 105,4 MPa in untreated wood, and 96 MPa and 113,5 MPa for the 220°C treatment under vacuum and nitrogen respectively. MOE was lower for the treatment under vacuum but had no significant differences under nitrogen (Table 5). For peroba mica (Table 5), the MOR increased with both treatments e.g. from 97 MPa in the untreated wood to 109 MPa and 125 MPa for vacuum and nitrogen at 220 °C, respectively. MOE increased considerably by 35% in relation to the untreated wood with the vacuum and nitrogen methods at 220 °C.
For cumaru wood (Table 4), no influence of temperature and of treatment method was observed. Reports in the literature regarding MOE and MOR changes with heat treatments are not always coincident. Poncsak (2006) reports that the wood becomes more rigid and brittle, and mechanical properties may be reduced, depending on temperature, heating rate and duration of treatment. An inert or slightly reducing atmosphere reduces the loss of mechanical strength (Doi et al. 1999).
A reduction in MOR has been reported and explained by the changes in content and structure of hemicelluloses induced by the heat treatment, causing loss of flexural strength of wood (Boonstra et al. 2007, Korkut et al. 2008. MOE increases in heat-treated wood, and Gunduz et al. (2009) explained this result by increased lignin cross-linking that makes the structure around the cellulose microfibrils and the middle lamella more rigid. Esteves and Pereira (2009) explained this increase in MOE with increasing crystallinity of the cellulose and the reduction of moisture content equilibrium.
Physical and mechanical properties ..: Araujo et al.
CONCLUSIONS
The heat treatment of bracatinga, peroba mica and cumaru wood significantly lowered the equilibrium moisture content, although with only a small impact on improving dimensional stability. A vacuum environment was more effective to reduce wood hygroscopicity.The heat-treated woods showed in general higher hardness and bending resistance. The wood from the different species showed a different response to the heat treatment, suggesting that optimization of the heat treatment operational parameters should be species' specific. The heat treatment intensified qualities of wood for use of the external constructions and floors. | 4,886 | 2016-01-01T00:00:00.000 | [
"Materials Science",
"Agricultural And Food Sciences"
] |
Low-Cost Online Partial Discharge Monitoring System for Power Transformers
The article presents in detail the construction of a low-cost, portable online PD monitoring system based on the acoustic emission (AE) technique. A highly sensitive piezoelectric transducer was used as the PD detector, whose frequency response characteristics were optimized to the frequency of AE waves generated by discharges in oil–paper insulation. The popular and inexpensive Teensy 3.2 development board featuring a 32-bit MK20DX256 microcontroller with the ARM Cortex-M4 core was used to count the AE pulses. The advantage of the system is its small dimensions and weight, easy and quick installation on the transformer tank, storage of measurement data on a memory card, battery power supply, and immediate readiness for operation without the need to configure. This system may contribute to promoting the idea of short-term (several days or weeks) PD monitoring, especially in developing countries where, with the dynamically growing demand for electricity, the need for inexpensive transformer diagnostics systems is also increasing. Another area of application is medium-power transformers (up to 100 MVA), where temporary PD monitoring using complex measurement systems requiring additional infrastructure (e.g., control cabinet, cable ducts for power supply, and data transmission) and qualified staff is economically unjustified.
Introduction
The results of the international survey on substation transformer failures published by CIGRE showed that, in most cases, they were caused by partial discharges [1,2]. Apart from the relatively rarely committed serious manufacturing errors (e.g., non-degassed or under-dried insulation system, the use of conductive structural elements such as screws or nuts with sharp edges, metal particles left after assembly works, incorrectly designed elements of the insulation system, etc.), partial discharges are initiated in these areas of the insulating system that are highly moist and degraded as a result of aging processes (pyrolysis, hydrolysis, and oxidation reactions) [3][4][5]. Partial discharges can also be generated in the vicinity of deformed windings due to overvoltage or mechanical shock during transformer transportation. For this reason, the phenomenon of partial discharges is not only the cause of failure but can be, and in fact is, more and more commonly treated as a reliable indicator of the condition of the transformer insulation system [6,7].
For PD detection in laboratory conditions, the conventional electrical method IEC 60270 [8] is used, while in the place of installation of the power transformer, the acoustic emission (AE) or ultra-high frequency (UHF) method is preferable due to the resistance to external electromagnetic interferences [9,10]. These unconventional techniques of PD detection have been adapted to both periodic diagnostic tests and tests carried out in the online monitoring mode [11][12][13]. The main advantage of monitoring over periodic diagnostics is the possibility of immediate detection of PD ignition or an increase in their intensity and, thus, allows operating services to prepare quickly and implement procedures minimizing the risk of failure. Based on the data collected by the monitoring system, one can dynamically change the requirements for periodic diagnostic tests (e.g., their acceleration or postponement) and decide if and when it is necessary to plan downtime and perform maintenance. Some online PD monitoring systems usually allow data exchange with a substation SCADA system. This enables the implementation of advanced inference rules based on statistical or machine learning methods, which allow detection of the relationships between the PD activity and such monitored parameters as voltage, load, oil temperature, tap position of the OLTP, or the amount of hydrogen dissolved in the oil [14][15][16].
In addition to the undoubted benefits of using online partial discharge monitoring systems, a serious obstacle to their dissemination is still a small number of specialized manufacturers, which translates into low supply and the high price of the device. In addition, the price is increased by the high costs of research and development and production. For this reason, only the largest transformers of key importance for the power system are currently equipped with online partial discharge monitoring systems. Therefore, in recent years, there has been an intensification of research work related to the minimization of the costs of PD monitoring. Two main areas of activity can be distinguished in the conducted research. The first one is related to the development of simple and low-cost partial discharge sensors. Castro et al. discussed [17] the possibility of using piezoelectric membranes (buzzers) to detect PD in power transformers based on the acoustic emission method. Piezoelectric membranes are readily available and often cheaper than conventional AE sensors, making their use particularly attractive in applications where several sensors are used. Laboratory tests have shown that a low-cost piezoelectric membrane provides acceptable partial discharge detection sensitivity and can be an alternative to expensive AE sensors. Besharatifard et al. [18] investigated the possibility of replacing piezoelectric technology with microfiber composites (MFCs). Compared to piezoelectric membranes, MFC sensors are slightly more expensive but offer higher sensitivity up to 500 kHz and a longer pulse duration, which can help separate partial discharges from on-site noise. For several years, also in the case of electromagnetic PD detectors operating in the radio frequency bands (HF, VHF, and UHF), research has been carried out to simplify their design and reduce production costs [19][20][21][22]. An example of such designs are miniaturized UHF antennas, which can be mass-produced using printed circuit board (PCB) technology [23]. According to the literature on the subject, the most promising designs of UHF PD detectors, which can be easily manufactured in PCB technology, are microstrip patch antennas [24][25][26], bio-inspired antennas [27,28], meander-line antennas [29], logarithmic spiral antennas [30], Vivaldi antennas [31], different types of Archimedean spiral antennas [32,33], meandered planar inverted-F antennas, and fractal antennas such as Hilbert curve fractal antennas [34,35], Peano fractal antennas [36], H-fractal antennas [37], or Minkowski fractal antennas [38]. In addition to UHF antennas, the most commonly used low-cost electromagnetic PD detectors are high-frequency current transformers (HFCT) and transient earth voltage (TEV) sensors. Their advantage is not only simple construction but also high sensitivity [39,40].
The second area of research concerns the construction of partial discharge monitoring systems based on generally available single-chip microcontrollers, simple USB data loggers, or field programmable gate arrays (FPGA), which are an alternative to expensive, multichannel data acquisition cards. Saeed et al. [41] presented a supervisory system for PD monitoring designed around the Microchip PIC24EP512GU810 microcontroller, which is characterized by a high speed of 70 million instructions per second (MIPS). The device also offers a satisfying analog-to-digital conversion performance of 1.1 million samples per second with up to four simultaneous channels. This ensures that the fast PD pulses are adequately sampled. Chakrabarty et al. [42] have developed an online partial discharge counting system using a microcontroller and FPGA technology. The signal from the PD detector is sampled at a frequency of 20 MHz by the analog-to-digital converter of the PIC 16F877A microcontroller. The signal is then transmitted to the input of a Xilinx ML405 evaluation board equipped with an FPGA chipset type Virtex-4 (XC4VFX20) that has been programmed to detect and count PD pulses in real time. Chang et al. [43] have developed a reconfigurable partial discharge monitoring system based on FPGA technology. The proposed system uses the AMR Cortex M4 microprocessor to control the system and data analysis, a fast AD9226 analog-to-digital converter, and Xilinx XC6SLX16 field programmable gate arrays with SDRAM memory. The signal processing procedure consists of two steps. In the first step, PD pulses are captured by various PD detectors, quickly converted to digital data by AD9226, and then temporarily stored in SDRAM by an FPGA chip. In the second step, the signals are taken from the SDRAM and transmitted to the microprocessor by the FPGA for further data analysis. Yan et al. [44] presented the PD monitoring and locating system for medium-voltage switchgears, which is based on low-cost TEV detectors. In order to automatically locate the PD source and minimize the number of expensive high-speed acquisition cards, the authors proposed a time-sharing access mechanism, which was implemented by multiple high-frequency surface mounting relays integrated into each TEV detector. Mohamed et al. [45] proposed the use of a spectrum analyzer based on the SDR (software-defined radio) technology to build a cheap and portable online PD monitoring system operating in the VHF/UHF frequency range. A system that is capable of automatically collecting PD data consists of only two hardware components, i.e., a PC/laptop and a portable software-defined radio receiver type Realtek RTL2832U that connects to the computer via a USB interface.
This article presents in detail the construction of a low-cost, portable PD online monitoring system based on the acoustic emission (AE) technique, which has a chance to fill the gap in the area of inexpensive monitoring systems that can be effectively used in the diagnosis of small and medium power transformers. Currently designed systems and commercial systems available on the market are intended-mainly due to the high price and complicated construction-to monitor large power transformers that play a strategic role in the power system. To the best of the authors' knowledge, this manuscript is the first complete description of a portable online PD monitoring system that can be assembled from commercially available electronics and development boards compatible with the Arduino architecture. As a module for processing and analyzing AE signals, the Teensy 3.2 board was used. Despite its small size and low price (20USD), it ensures high efficiency of detection and counting of AE pulses generated by partial discharges. Laboratory tests have shown that the system is capable of lossless, real-time counting of up to over 83,000 AE pulses.
The developed device is the world's first PD monitoring system equipped with a piezoelectric acoustic emission sensor optimized for the detection of partial discharges in oil-paper insulation. The frequency characteristics and resonant frequencies of the sensor coincide with the frequencies of AE pulses generated by the most dangerous and destructive types of discharges for the transformer insulation system, i.e., inter-turn discharges, creeping and surface discharges. As a result, the presented low-cost monitoring system has a very high sensitivity of PD detection. What is also essential is that the system is in line with the strategy of modern monitoring of power transformers formulated in the Cigre TB 630 brochure published in 2015: Guide On Transformer Intelligent Condition Monitoring (TICM) Systems, which assumes the replacement of traditional sensors and transducers with Intelligent Electronic Devices (IED), which are equipped with a microprocessor and are capable of processing raw PD pulses and automatically determining the parameters describing them.
The developed system may contribute to the promotion of the idea of short-term (several days or weeks) PD monitoring, especially in developing countries where, with the dynamically growing demand for electricity, the demand for inexpensive transformer diagnostics systems is also growing. Another area of application is medium power transformers (up to 100 MVA), on which temporary PD monitoring using expensive, often requiring additional infrastructure (e.g., control cabinet, cable ducts for power and data transmission), and qualified staff would be economically unjustified.
This paper is organized as follows. The hardware and software layers of the system are discussed in Section 2, while Section 3 presents the results of laboratory and field tests that allowed us to assess the possibility of using the system for short-term PD monitoring
General Description
In designing the device, the following assumptions were made: low price and availability of all system components, easy assembly, high sensitivity of PD detection, resistance to external electromagnetic and acoustic disturbances, registration of the PD data in real time, collecting PD data and storing it on a portable, inexpensive storage medium, battery or mains operation, small dimensions and weight, and immediate readiness to work.
The online PD monitoring system is based on the acoustic emission method, which, along with electromagnetic HF/VHF/UHF methods, belongs to the group of unconventional partial discharge detection techniques. The main advantages of the AE method include relatively high sensitivity, which primarily depends on the position of the sensor in relation to the PD source, galvanic separation of the tested object from the measurement system, resistance to external electromagnetic interference, the ability to locate PD using auscultation or TDOA (time difference of arrival) technique, and installation of the measurement system does not require switching off the tested object. As shown in the schematic diagram of the system (Figure 1), a contact piezoelectric transducer was used to detect acoustic waves from PD. The electrical signals at the output of the piezoelectric transducer usually have very small amplitude, from a few to tens of millivolts, therefore, the monitoring system is equipped with a preamplifier circuit consisting of an instrumentation amplifier with adjustable gain (default gain is 40 dB), and a voltage follower. Another element of the system is an active bandpass filter, whose task is to eliminate low-frequency components of the acoustic background (e.g., oil pump and cooling fan noises, vibrations caused by the magnetostrictive action of the transformer core, etc.) and high-frequency electromagnetic interferences. The amplified and filtered signal is then processed by peak detector and voltage comparator circuits. The use of both of these systems made it possible to replace the expensive signal acquisition card with a simple microcontroller equipped with an analog-to-digital converter. The operation of the monitoring system is controlled by a program written in C/C++, which performs such functions as detection and counting PD pulses, saving measurement data on a memory card, and presenting current measurement data on the LCD.
Partial Discharge Detector
A contact piezoelectric transducer was used as a partial discharge detector and the various design and production stages, described in detail in Reference [46]. The frequency response of the transducer was determined on the basis of laboratory tests aimed at identifying the acoustic frequencies emitted by partial discharges in oil-paper insulation. The test results showed that the most destructive forms of PD for the transformer insulation system (inter-turn, surface, and creeping discharges) generate AE signals, the energy of
Partial Discharge Detector
A contact piezoelectric transducer was used as a partial discharge detector and the various design and production stages, described in detail in Reference [46]. The frequency response of the transducer was determined on the basis of laboratory tests aimed at identifying the acoustic frequencies emitted by partial discharges in oil-paper insulation. The test results showed that the most destructive forms of PD for the transformer insulation system (inter-turn, surface, and creeping discharges) generate AE signals, the energy of which is transferred in three bands: 20-45 kHz, 50-70 kHz, and 85-115 kHz. The dominant frequencies in these bands are 40 kHz, 68 kHz, and 90 kHz, respectively. Since Barkhausen noise from the transformer core can reach ultrasonic frequencies [47], the transducer has been optimized to work in the other two higher frequency bands characteristic for partial discharges, i.e., 50-70 kHz and 85-115 kHz. The optimal material and geometric properties of the main transducer structures were selected using the Krimholtz-Leedom-Matthaei (KLM) model. Based on the simulation results, two piezoelectric disks made of PZT-5A (Navy Type II) ceramics with a diameter of 10.5 mm and a height of 18.8 mm and 25 mm, respectively, were used to build the transducer. Piezoelectric elements with opposite polarization directions were bonded to the matching layer using an electrically conductive composite adhesive based on epoxy resin and silver. The matching layer is made of a round plate with a diameter of 25 mm and a height of 1 mm, made of high-density alumina ( Figure 2a). The acoustic impedance of the matching layer is 37.9 MRayl, which ensures an efficient transfer of acoustic wave energy from the mineral oil through the steel tank of the power transformer to the piezoelectric elements of the transducer. Everything is closed in a housing made of stainless steel. The transducer was placed centrally on the front wall of the monitoring system housing, which was additionally equipped with four height-adjustable magnetic holders (Figure 2b).
The selected geometric and material parameters, as well as a fully differential design, allowed to obtain the desired properties of the transducer, i.e., a two-resonant (68 kHz and 90 kHz) and wide (30-100 kHz) frequency response curve and high peak sensitivity (−61.1 dB ref. V/µbar) (Figure 2c). The test results presented in reference [46] showed that this transducer is characterized by high detection sensitivity of partial discharges generated in paper-oil insulation. Compared to commonly used commercial AE sensors, the average PD pulse amplitude recorded by the new transducer was a minimum of 5.2 dB higher and a maximum of 19.8 dB higher. Figure 2d shows the AE waveforms recorded a needle partial discharge in oil with an apparent charge of 82 pC and a surface discharge on a pressboard sample in oil with an apparent charge of 387 pC. Figure 3 shows circuit diagrams and photographs of the amplifier and filter designed for the PD monitoring system. The amplifier circuit is based on the AD8421BRZ instrumentation amplifier from Analog Devices (Analog Devices, Norwood, MA, USA). The gain G is regulated by the appropriate selection of the resistance R G value. In this case, a 100-ohm resistor was used, giving a gain of 40 dB (G = 100). The main advantages of this circuit are low current consumption (<2.3 mA), low noise level, which does not exceed 3.2 nV/ √ Hz, ultra-low polarization current (<500 pA), wide bandwidth (2 MHz with gain G = 100) and low price (around $10).
Amplifier and Bandpass Filter
The active bandpass filter (20-500 kHz) is designed based on the Sallen-Key architecture, where the low-pass and high-pass sections have a fourth-order Butterworth filter structure with unity gain. For this purpose, the Analog Devices ADA4898 voltage feedback operational amplifier was used, which is characterized by ultralow noise (0.9 nV/ √ Hz), ultralow distortion (−93 dBc at 500 kHz), and low supply current (8 mA). The filter is made on a separate PCB and is plugged in using dedicated board-to-board connectors between the instrumentation amplifier and the voltage follower. This solution allows the filter to be easily disconnected or replaced with a filter with a different passband characteristic.
The designed low-cost PD monitoring system provides the possibility of transmitting the amplified and filtered AE signal via a coaxial cable directly to an external measuring device (oscilloscope, signal acquisition card) or to a substation SCADA system. Coaxial cables have a relatively large capacity (50-100 pF/m) and, therefore, cannot be directly connected to the output of the amplifier. This problem was solved by the use of a voltage follower based on the AD810 op-amp, which can drive capacitive loads exceeding 1000 pF, without parasitic oscillations. Figure 3 shows circuit diagrams and photographs of the amplifier and filter designed for the PD monitoring system. The amplifier circuit is based on the AD8421BRZ instrumentation amplifier from Analog Devices (Analog Devices, Norwood, MA, USA). The gain G is regulated by the appropriate selection of the resistance RG value. In this case, a 100-ohm resistor was used, giving a gain of 40 dB (G = 100). The main advantages of this circuit are low current consumption (<2.3 mA), low noise level, which does not exceed 3.2 nV/√Hz, ultra-low polarization current (<500 pA), wide bandwidth (2 MHz with gain G = 100) and low price (around $10).
Amplifier and Bandpass Filter
The active bandpass filter (20-500 kHz) is designed based on the Sallen-Key architecture, where the low-pass and high-pass sections have a fourth-order Butterworth filter structure with unity gain. For this purpose, the Analog Devices ADA4898 voltage feedback operational amplifier was used, which is characterized by ultralow noise (0.9 nV/√Hz), ultralow distortion (−93 dBc at 500 kHz), and low supply current (8 mA). The filter is made on a separate PCB and is plugged in using dedicated board-to-board con-
Peak Detector and Voltage Comparator
In order to reduce the required sampling frequency of the AE pulses, a passive peak detector (sometimes called an envelope detector) and a voltage comparator were used. The peak detector is made up of just three components: fast-switching diode D type 1N4148, capacitor C with capacitance 22 pF, and resistor R with resistance 1 MΩ (Figure 4a). Together, these elements form a half-wave rectifier that charges the capacitor C to the peak voltage of the incoming AE burst. As the amplitude of the input signal increases, the capacitor voltage is increased by a rectifier diode D. When the amplitude of the input signal decreases, the capacitor is discharged through a parallel bleeder resistor R. The discharge rate of the capacitor depends on the value of the time constant τ = RC. During the first time constant, the capacitor discharges 63%, and after 5τ, it is almost completely discharged. The designed low-cost PD monitoring system provides the possibility of transmitting the amplified and filtered AE signal via a coaxial cable directly to an external measuring device (oscilloscope, signal acquisition card) or to a substation SCADA system. Coaxial cables have a relatively large capacity (50-100 pF/m) and, therefore, cannot be directly connected to the output of the amplifier. This problem was solved by the use of a voltage follower based on the AD810 op-amp, which can drive capacitive loads exceeding 1000 pF, without parasitic oscillations.
Peak Detector and Voltage Comparator
In order to reduce the required sampling frequency of the AE pulses, a passive peak detector (sometimes called an envelope detector) and a voltage comparator were used. The peak detector is made up of just three components: fast-switching diode D type 1N4148, capacitor C with capacitance 22 pF, and resistor R with resistance 1 MΩ ( Figure 4a). Together, these elements form a half-wave rectifier that charges the capacitor C to the peak voltage of the incoming AE burst. As the amplitude of the input signal increases, the capacitor voltage is increased by a rectifier diode D. When the amplitude of the input signal decreases, the capacitor is discharged through a parallel bleeder resistor R. The discharge rate of the capacitor depends on the value of the time constant τ = RC. During the first time constant, the capacitor discharges 63%, and after 5τ, it is almost completely discharged.
A voltage comparator is a special type of operational amplifier with an unbalanced input and high gain that compares the voltage value of the signal applied to the noninverting input with the reference voltage applied to the inverting input. The voltage value at the comparator's output results from the difference in voltages between its two inputs. If the voltage at the non-inverting input is higher than at the inverting input, then the output voltage is close to the positive pole of the supply. If the voltage at the noninverting input is lower than at the inverting input, then the output voltage is close to the negative pole of the power supply. Thus, the comparator can be considered an elementary, one-bit analog-to-digital converter. The project uses a single-channel LM311N voltage comparator (Figure 4b), which is characterized by a very low current consumption (up to 100 nA at an ambient temperature of 25 °C) and a fast response time (~200 ns). The open collector (OC) comparator output is compatible with CMOS, TTL, and RTL-DTL integrated circuits and can switch voltages up to 50 V and currents up to 50 mA. The LM311N is designed to operate with a wide range of supply voltages, including ±15 V power supplies for operational amplifiers and, as in this case, 5 V for logic circuits. The voltage value given to the inverting input is set with the A20k rotary potentiometer. The correctness of the peak detector and voltage comparator operation was tested in a measurement set-up shown in Figure 5a. As the AE signal source, the Olympus V101B ultrasonic probe was used, which was acoustically coupled to a piezoelectric transducer with a USG gel. The ultrasonic probe was excited by rectangular pulses with a voltage of 10 to 500 mV and a duration of 1 μs to 3 μs, which were generated using a Keysight DSOX2024A oscilloscope with a built-in waveform generator. The frequency with which the pulses were generated varied in the range from 1 Hz to 1 kHz. The outputs of the voltage follower, peak detector, and voltage comparator were connected with coaxial ca- at the comparator's output results from the difference in voltages between its two inputs. If the voltage at the non-inverting input is higher than at the inverting input, then the output voltage is close to the positive pole of the supply. If the voltage at the non-inverting input is lower than at the inverting input, then the output voltage is close to the negative pole of the power supply. Thus, the comparator can be considered an elementary, one-bit analog-to-digital converter. The project uses a single-channel LM311N voltage comparator (Figure 4b), which is characterized by a very low current consumption (up to 100 nA at an ambient temperature of 25 • C) and a fast response time (~200 ns). The open collector (OC) comparator output is compatible with CMOS, TTL, and RTL-DTL integrated circuits and can switch voltages up to 50 V and currents up to 50 mA. The LM311N is designed to operate with a wide range of supply voltages, including ±15 V power supplies for operational amplifiers and, as in this case, 5 V for logic circuits. The voltage value given to the inverting input is set with the A20k rotary potentiometer.
The correctness of the peak detector and voltage comparator operation was tested in a measurement set-up shown in Figure 5a. As the AE signal source, the Olympus V101B ultrasonic probe was used, which was acoustically coupled to a piezoelectric transducer with a USG gel. The ultrasonic probe was excited by rectangular pulses with a voltage of 10 to 500 mV and a duration of 1 µs to 3 µs, which were generated using a Keysight DSOX2024A oscilloscope with a built-in waveform generator. The frequency with which the pulses were generated varied in the range from 1 Hz to 1 kHz. The outputs of the voltage follower, peak detector, and voltage comparator were connected with coaxial cables to channels 1, 2, and 3 of the oscilloscope, respectively. Examples of registered waveforms are shown in Figure 5b. In this case, the ultrasonic probe was excited by rectangular pulses generated with a frequency of 1 kHz, with a duration of 3 µs, and an amplitude of 50 mV.
Single Board Microcontroller
The popular and inexpensive ($20) Teensy 3.2 development board featuring a 32-bit MK20DX256 microcontroller with the ARM Cortex-M4 core was used to count the PD pulses. The Teensy 3.2 module, despite its small dimensions (35 × 18 mm), has 34 digital input/output lines tolerant of 5V voltage, of which 12 can be used as PWM outputs, as well as one analog output. Between them, there are 21 high-resolution analog inputs (13 usable bits), of which 16 lines are shared with the digital ones and one with the analog output line (they cannot be used simultaneously). In addition, the board is equipped with 7 timers, 3 UART serial ports, SPI, I2C, I2S, CAN Bus, RTC module, 16 DMA channels, and touch sensor inputs.
The output signal from the voltage comparator is fed directly to Pin 13 of the microcontroller. If the amplitude of the signal at the input of the comparator is higher than the set trigger level, then its output returns a logic "high" state (~3.3 V). Otherwise, a logic "low" state (~0 V) is sent to Pin 13. The microcontroller continuously counts the PD pulses and records their current number every minute on the microSD memory card (Open Smart Technologies Limited, Hong Kong, China) and displays it on the 8 × 2 character monochrome LCD display type WC0802C (Hubbell Wiegmann, Freeburg, IL, USA). The schematic diagram of connecting the LCD driver and the microSD card adapter to the Teensy 3.2 board is shown in Figure 6.
Power Supply
The monitoring system is powered by a package of four series-connected lithiumion cells, type US18650VTC5A (Murata Manufacturing Co., Nagaokakyo, Japan), with a nominal voltage of 3.6 V and a capacity of 2600 mAh, which cooperate with a dedicated module of the BMS 4S battery management system (Figure 7a). The module has the function of charging cells with the option of a balancer, discharging, and a function protecting the cells against excessive discharge. The cells can be charged with a continuous current of up to 10 A. Since all digital modules (microcontroller, LCD driver, memory card adapter) require a voltage of 5 V, behind the battery management system, there is a step-down pulse converter module whose output voltage value is set with a potentiometer (from 1.0 V to 17 V). In turn, the instrumentation amplifier, voltage follower, and operational amplifiers used to build an active bandpass filter require a symmetrical voltage of 15 V. For this reason, the power module is additionally equipped with a voltage converter type ICL7662CPA+ (Figure 7b). This converter is a monolithic charge pump voltage inverter that converts a positive voltage in the range of +4.5 V to +20 V to a corresponding negative voltage of −4.5 V to −20 V. set trigger level, then its output returns a logic "high" state (~3.3 V). Otherwise, a logic "low" state (~0 V) is sent to Pin 13. The microcontroller continuously counts the PD pulses and records their current number every minute on the microSD memory card (Open Smart Technologies Limited, Hong Kong, China) and displays it on the 8 × 2 character monochrome LCD display type WC0802C (Hubbell Wiegmann, Freeburg, IL, USA). The schematic diagram of connecting the LCD driver and the microSD card adapter to the Teensy 3.2 board is shown in Figure 6.
Power Supply
The monitoring system is powered by a package of four series-connected lithium-ion cells, type US18650VTC5A (Murata Manufacturing Co., Nagaokakyo, Japan), with a nominal voltage of 3.6 V and a capacity of 2600 mAh, which cooperate with a dedicated module of the BMS 4S battery management system (Figure 7a). The module has the function of charging cells with the option of a balancer, discharging, and a function protecting the cells against excessive discharge. The cells can be charged with a continuous current of up to 10 A. Since all digital modules (microcontroller, LCD driver, memory card adapter) require a voltage of 5 V, behind the battery management system, there is a step-down pulse converter module whose output voltage value is set with a potentiometer (from 1.0 V to 17 V). In turn, the instrumentation amplifier, voltage follower, and operational amplifiers used to build an active bandpass filter require a symmetrical voltage of 15 V. For this reason, the power module is additionally equipped with a voltage converter type ICL7662CPA+ (Figure 7b). This converter is a monolithic charge pump voltage inverter that converts a positive voltage in the range of +4.5 V to +20 V to a corresponding negative voltage of −4.5 V to −20 V.
Assembly of the System
All electronic modules of the monitoring system together with the power supply system are placed in a waterproof (IP67) aluminum housing with dimensions of 120 × 120 × 75 mm. On the front panel of the system, there is an LCD screen and a switch for resetting the pulse counter. The front panel elements are protected by a transparent inspection window made of polycarbonate. In addition, the housing is equipped with a DC power connector through which the charger is connected, a micro SD card slot, a grounding screw, and a BNC connector for optional wired AE signal transmission (Figure 8).
Assembly of the System
All electronic modules of the monitoring system together with the power supply system are placed in a waterproof (IP67) aluminum housing with dimensions of 120 × 120 × 75 mm. On the front panel of the system, there is an LCD screen and a switch for resetting the pulse counter. The front panel elements are protected by a transparent inspection window made of polycarbonate. In addition, the housing is equipped with a DC power connector through which the charger is connected, a micro SD card slot, a grounding screw, and a BNC connector for optional wired AE signal transmission (Figure 8).
Firmware
Each Teensy 3.2 module has a bootloader uploaded, thanks to which it can be programmed using the built-in USB connector (no external programmer required). Programs can be written in any environment supporting the C language or-as in the case of the discussed monitoring system-in the Arduino IDE programming environment with the Teensyduino extension installed. The computer program performs three main functions, i.e., counting PD pulses, displaying the number of pulses on the LCD display, and storing data (time, date, and number of pulses) on the memory card. The complete source code with comments can be found in the Appendix A.
tem are placed in a waterproof (IP67) aluminum housing with dimensions of 120 × 120 × 75 mm. On the front panel of the system, there is an LCD screen and a switch for resetting the pulse counter. The front panel elements are protected by a transparent inspection window made of polycarbonate. In addition, the housing is equipped with a DC power connector through which the charger is connected, a micro SD card slot, a grounding screw, and a BNC connector for optional wired AE signal transmission (Figure 8).
System Testing
The test of the partial discharge monitoring system was carried out in three stages. In the first stage, the efficiency of the AE pulse counting function was checked using a signal generator and a piezoelectric transmitter. The second stage of the test was carried out in controlled laboratory conditions, during which partial discharges generated in the transformer tank model were monitored. The third stage of the test was carried out in field conditions, and the test object was a 10 MVA power transformer.
Testing the Efficiency of AE Pulse Counting
To assess the efficiency of counting AE pulses, a measurement system was used consisting of an oscilloscope Keysight with a built-in signal generator, a reference pulse counter, and a wideband piezoelectric transducer Olympus V101B, which was excited with rectangular pulses with a duration of 1 us and an amplitude of 100 mV (Figure 9). can be written in any environment supporting the C language or-as in the case of the discussed monitoring system-in the Arduino IDE programming environment with the Teensyduino extension installed. The computer program performs three main functions, i.e., counting PD pulses, displaying the number of pulses on the LCD display, and storing data (time, date, and number of pulses) on the memory card. The complete source code with comments can be found in the Appendix A.
System Testing
The test of the partial discharge monitoring system was carried out in three stages. In the first stage, the efficiency of the AE pulse counting function was checked using a signal generator and a piezoelectric transmitter. The second stage of the test was carried out in controlled laboratory conditions, during which partial discharges generated in the transformer tank model were monitored. The third stage of the test was carried out in field conditions, and the test object was a 10 MVA power transformer.
Testing the Efficiency of AE Pulse Counting
To assess the efficiency of counting AE pulses, a measurement system was used consisting of an oscilloscope Keysight with a built-in signal generator, a reference pulse counter, and a wideband piezoelectric transducer Olympus V101B, which was excited with rectangular pulses with a duration of 1 us and an amplitude of 100 mV (Figure 9). The measurement results listed in Table 1 show that the system is characterized by high efficiency and is capable of lossless counting of up to 83,400 AE pulses per minute. Above this value, the counting accuracy decreases. The measurement results listed in Table 1 show that the system is characterized by high efficiency and is capable of lossless counting of up to 83,400 AE pulses per minute. Above this value, the counting accuracy decreases.
Online Monitoring of Partial Discharges in Laboratory Conditions
The tests were carried out in a shielded high-voltage laboratory using a model of a power transformer tank with dimensions of 1200 × 800 × 730 mm, which was filled with mineral oil. A system of electrodes for generating surface discharges on a round sample of pressboard with a diameter of 100 mm and a thickness of 3 mm was mounted to the internal connecting terminal of the bushing. The tip of a point electrode was located 400 mm from the wall of the ladle on which the tested system was installed. The PDtracker Portable system (Poznan University of Technology, Poznan, Poland) was used as a reference system, which is equipped with eight analog inputs with parallel sampling up to 20 MS/s. As partial discharge detectors, the system accepts both piezoelectric acoustic emission sensors (default configuration) and high-frequency current transformers. The AE sensor of the tested system and the A6890 sensor of the reference PDtracker system were placed exactly at the height of the PD source. The distance between the sensors was 200 mm, while the distance from the PD source to each sensor was the same and was about 400 mm ( Figure 10). One of the basic functions performed by the PDtracker Portable system is also the counting of PD pulses. For this purpose, the system continuously records the signal, which is divided into time frames with a fixed width set by the user (in this case, 2 ms). If, in the analyzed time frame, the amplitude of the AE waveform exceeds the threshold value, then the system treats this event as the occurrence of a PD pulse and counts it. The threshold value is usually twice the average amplitude of the acoustic background noise, which in laboratory conditions does not exceed 10 mV.
During the tests, in addition to the reference PDtracker Portable system discussed above, a conventional PD meter (PD-Smart, Doble Engineering Company, Marlborough, MA, USA) and a measuring circuit compliant with the IEC60270 standard were also used ( Figure 11). This made it possible to control the value of the apparent charge of the partial discharges and their intensity during the experiment.
A voltage in the range from 0 to 17 kV was applied to the electrode system, which allowed to generate partial discharge pulses of various intensity and energy. The inception voltage of surface discharges was U i = 9.2 kV, and their apparent charge ranged from 45 pC to over 5 nC. Exemplary test results are shown in Figure 12.
The analysis of the obtained test results shows that the discussed system is capable of online monitoring of partial discharges occurring in the oil-paper insulation system of the transformer. Due to the different implementations of the pulse counting procedure by the tested and reference acoustic emission systems, correlation analysis was performed instead of quantitative analysis. The Pearson correlation coefficient determined for the number of AE pulses recorded by both systems was 0.775 (strong correlation), while Spearman's correlation coefficient was 0.91 (very strong correlation). In turn, the comparison with the number of PD pulses recorded by a conventional meter was more favorable for the PDtracker Portable reference system, as in this case, the Pearson correlation coefficient was 0.797, while for the low-cost monitoring system it was 0.639, which means a moderate correlation. Spearman's correlation coefficient was 0.891 and 0.793, respectively, which can be interpreted as a strong correlation.
placed exactly at the height of the PD source. The distance between the sensors was mm, while the distance from the PD source to each sensor was the same and was a 400 mm (Figure 10). One of the basic functions performed by the PDtracker Portable tem is also the counting of PD pulses. For this purpose, the system continuously rec the signal, which is divided into time frames with a fixed width set by the user (in case, 2 ms). If, in the analyzed time frame, the amplitude of the AE waveform exceed threshold value, then the system treats this event as the occurrence of a PD pulse counts it. The threshold value is usually twice the average amplitude of the acoustic b ground noise, which in laboratory conditions does not exceed 10 mV.
(a) (b) Figure 10. Test of the monitoring system in laboratory conditions: (a) electrode system for gen ing surface discharges; (b) arrangement of the tested and reference monitoring system on the t former tank model: TT-oil filled transformer tank with electrode system for generating partia charges; TB-transformer bushing; PD-electrode system for generating surface partial disch PT-piezoelectric transducer A6890; DUT-system under test.
During the tests, in addition to the reference PDtracker Portable system discu above, a conventional PD meter (PD-Smart, Doble Engineering Company, Marlboro MA, USA) and a measuring circuit compliant with the IEC60270 standard were also ( Figure 11). This made it possible to control the value of the apparent charge of the pa discharges and their intensity during the experiment.
A voltage in the range from 0 to 17 kV was applied to the electrode system, w allowed to generate partial discharge pulses of various intensity and energy. The in tion voltage of surface discharges was Ui = 9.2 kV, and their apparent charge ranged 45 pC to over 5 nC. Exemplary test results are shown in Figure 12. Figure 10. Test of the monitoring system in laboratory conditions: (a) electrode system for generating surface discharges; (b) arrangement of the tested and reference monitoring system on the transformer tank model: TT-oil filled transformer tank with electrode system for generating partial discharges; TB-transformer bushing; PD-electrode system for generating surface partial discharges PT-piezoelectric transducer A6890; DUT-system under test.
Sensors 2023, 23, x FOR PEER REVIEW 14 of 21 Figure 11. The measuring set-up used during testing of the monitoring system in laboratory conditions: U-high-voltage supply; Z-short-circuit current limiting resistor; MS-reference online PD monitoring system PDtracker; PT-piezoelectric transducer A6890; DUT-system under test; TToil filled transformer tank with electrode system for generating partial discharges; CK-coupling capacitor; CD-measuring impedance; CC-connecting cable; M-conventional partial discharge measuring device Doble PD-Smart; PC-computer. Figure 11. The measuring set-up used during testing of the monitoring system in laboratory conditions: U-high-voltage supply; Z-short-circuit current limiting resistor; MS-reference online PD monitoring system PDtracker; PT-piezoelectric transducer A6890; DUT-system under test; TT-oil filled transformer tank with electrode system for generating partial discharges; CK-coupling capacitor; CD-measuring impedance; CC-connecting cable; M-conventional partial discharge measuring device Doble PD-Smart; PC-computer.
Online Monitoring of Partial Discharges in a 10 MVA Power Transformer
The field test of the system took place during partial discharge monitoring on a power transformer with a voltage of 115,000 ± 10% kV and a power of 10 MVA, which was manufactured in 1992. The reference measuring device was the eight-channel PDtracker Portable monitoring system, whose six piezoelectric transducers were placed near the phases of the low (LV1, LV2, LV3) and high (HV1, HV2, HV3) voltage sides. The last two transducers were installed on the side walls of the transformer tank (under the oil conservator and opposite the on-load tap-changer). The tested, low-cost PD monitoring system was always mounted in close proximity (approx. 15-20 cm) to the piezoelectric transducer of the reference system ( Figure 13). Figure 11. The measuring set-up used during testing of the monitoring system in laboratory conditions: U-high-voltage supply; Z-short-circuit current limiting resistor; MS-reference online PD monitoring system PDtracker; PT-piezoelectric transducer A6890; DUT-system under test; TToil filled transformer tank with electrode system for generating partial discharges; CK-coupling capacitor; CD-measuring impedance; CC-connecting cable; M-conventional partial discharge measuring device Doble PD-Smart; PC-computer.
Figure 12.
Test voltage value (a), PD apparent charge (b), number of PD pulses recorded by a conventional partial discharge meter PD-Smart, (c) number of AE pulses recorded by the reference monitoring system PDtracker Portable, (d) number of AE pulses registered by the tested, low-cost PD monitoring system (e). The analysis of the obtained test results shows that the discussed system is capable of online monitoring of partial discharges occurring in the oil-paper insulation system of the transformer. Due to the different implementations of the pulse counting procedure by the tested and reference acoustic emission systems, correlation analysis was performed instead of quantitative analysis. The Pearson correlation coefficient determined for the number of AE pulses recorded by both systems was 0.775 (strong correlation), while Spearman's correlation coefficient was 0.91 (very strong correlation). In turn, the comparison with the number of PD pulses recorded by a conventional meter was more favorable for the PDtracker Portable reference system, as in this case, the Pearson correlation coefficient was 0.797, while for the low-cost monitoring system it was 0.639, which means a moderate correlation. Spearman's correlation coefficient was 0.891 and 0.793, respectively, which can be interpreted as a strong correlation.
Online Monitoring of Partial Discharges in a 10 MVA Power Transformer
The field test of the system took place during partial discharge monitoring on a power transformer with a voltage of 115,000 ± 10% kV and a power of 10 MVA, which was manufactured in 1992. The reference measuring device was the eight-channel PDtracker Portable monitoring system, whose six piezoelectric transducers were placed near the phases of the low (LV1, LV2, LV3) and high (HV1, HV2, HV3) voltage sides. The last two transducers were installed on the side walls of the transformer tank (under the oil conservator and opposite the on-load tap-changer). The tested, low-cost PD monitoring system was always mounted in close proximity (approx. 15-20 cm) to the piezoelectric transducer of the reference system ( Figure 13). The measurement data analysis showed that in the tested transformer, both monitoring systems recorded a large number of AE pulses only near the HV3 phase ( Figure 14). The period of increased intensity of the acoustic emission phenomenon lasted about 20 min and began at the moment when the OLTC (on-load tap changer) position changed. After another tap change, both systems recorded only single AE events. The Pearson cor- Figure 13. Monitoring of partial discharges in the HV3 phase of a 10 MVA power transformer: MS-reference monitoring system PDtracker Portable; PT-piezoelectric transducer of the reference monitoring system; DUT-tested low-cost PD monitoring system.
The measurement data analysis showed that in the tested transformer, both monitoring systems recorded a large number of AE pulses only near the HV3 phase ( Figure 14). The period of increased intensity of the acoustic emission phenomenon lasted about 20 min and began at the moment when the OLTC (on-load tap changer) position changed. After another tap change, both systems recorded only single AE events. The Pearson correlation coefficient calculated for the number of AE pulses detected by both systems was 0.980 (very strong correlation), and the Spearman correlation coefficient was 0.869 (strong correlation). This showed that the performance of a low-cost PD monitoring system could be comparable to much more expensive and complex commercial systems.
Conclusions
The article presents in detail the design of a low-cost, portable online partial discharge monitoring system based on a non-invasive method of acoustic emission, which meets the guidelines of the IEC TS 62478 standard in terms of general requirements for the AE measurement system and for the measured PD quantities. The reason for developing this device was to eliminate the main obstacle to the widespread use of PD online monitoring systems, which is their high price and complicated operation, which requires a lot of experience. In the case of the discussed system, the total cost of all components used to build it does not exceed 300USD, which is a small fraction of the price of a commercial system. Thanks to this, the system can contribute to the dissemination of the idea of short-term PD monitoring, especially in developing countries, where with the dynamically growing demand for electricity, the demand for inexpensive, easy-to-manufacture transformer diagnostic systems is also growing. This system can also be used during laboratory measurements as a supplement or alternative to a stationary partial discharge detector. Another area of application for the device may be medium power transformers, for which online PD monitoring using complex measurement systems requiring additional infrastructure and qualified personnel is usually economically unjustified.
In addition to the low production price, the advantage of the system is: • a high detection sensitivity of acoustic signals from partial discharges, resulting from the use of true differential AE sensor with optimized frequency response characteristics; • the possibility of equipping the system with inexpensive modules dedicated to the Arduino/Teensyduino platform that increase its functionality, such as a Bluetooth module for wireless transmission of measurement data, or additional sensors to monitor other parameters of the transformer's operation (e.g., an accelerometer for measuring transformer tank vibrations); • easy installation on the transformer tank (thanks to the low weight and magnetic holders); • no need to configure the device; • high resistance to external electromagnetic interference due to the shielding of electronic modules and piezoelectric transducer elements.
Conclusions
The article presents in detail the design of a low-cost, portable online partial discharge monitoring system based on a non-invasive method of acoustic emission, which meets the guidelines of the IEC TS 62478 standard in terms of general requirements for the AE measurement system and for the measured PD quantities. The reason for developing this device was to eliminate the main obstacle to the widespread use of PD online monitoring systems, which is their high price and complicated operation, which requires a lot of experience. In the case of the discussed system, the total cost of all components used to build it does not exceed 300USD, which is a small fraction of the price of a commercial system. Thanks to this, the system can contribute to the dissemination of the idea of short-term PD monitoring, especially in developing countries, where with the dynamically growing demand for electricity, the demand for inexpensive, easy-to-manufacture transformer diagnostic systems is also growing. This system can also be used during laboratory measurements as a supplement or alternative to a stationary partial discharge detector. Another area of application for the device may be medium power transformers, for which online PD monitoring using complex measurement systems requiring additional infrastructure and qualified personnel is usually economically unjustified.
In addition to the low production price, the advantage of the system is: • a high detection sensitivity of acoustic signals from partial discharges, resulting from the use of true differential AE sensor with optimized frequency response characteristics; • the possibility of equipping the system with inexpensive modules dedicated to the Arduino/Teensyduino platform that increase its functionality, such as a Bluetooth module for wireless transmission of measurement data, or additional sensors to monitor other parameters of the transformer's operation (e.g., an accelerometer for measuring transformer tank vibrations); • easy installation on the transformer tank (thanks to the low weight and magnetic holders); • no need to configure the device; • high resistance to external electromagnetic interference due to the shielding of electronic modules and piezoelectric transducer elements.
The system also has some limitations, the most important of which is the battery power supply and the relatively short, several-day working time. Another disadvantage is related to the PD detection method used. In the case of the acoustic emission method, the sensitivity of partial discharge detection strongly depends on the distance between the AE sensor and the PD source. Therefore, before assembling the system, especially on large power transformers, it may be necessary to locate the place on the tank where the AE pulses are recorded in advance. A standard auscultatory technique (SAT) can be used for this purpose [48]. The possibility of real-time execution of only one task (counting PD pulses) can also be considered a disadvantage of the system. This is due to the limited computing power of the Teensy 3.2 microcontroller, which ranges from a few to several dozen MIPS. However, this problem can be relatively easily solved by equipping the system with a second microcontroller or using-unfortunately at the cost of many times more power consumption-a microcontroller with a much more efficient processor, such as Teensy 4.1 with an ARM Cortex-M7 processor clocked at 600 MHz. , en, , , d4, , , d5, , , d6, , , d7) //overwrite the number of pulses prior_count = count; } } } } } } //print the time on the LCD lcd.setCursor( ( (0, , , 0); ); ); printDigitsLcd( ( (hour()); ()); ()); lcd.print( ( (":"); ); ); printDigitsLcd( ( (minute()); ()); ()); lcd.print( ( (":"); ); ); printDigitsLcd( ( (second()); ()); ()); //save number of pulses to file in 1 minute intervals (seconds==0) //bool x is to avoid saving more | 12,175.8 | 2023-03-23T00:00:00.000 | [
"Physics"
] |
Cosmic evolution of dark energy in a generalized Rastall gravity
In this work, we propose a scheme for cosmic evolution in a generalized Rastall gravity. In our approach, the role of dark energy is taken by the non-conserved sector of the stress energy–momentum tensor. The resultant cosmic evolution is found to naturally consists of three stages, namely, radiation dominated, ordinary matter dominated, as well as dark energy and dark matter dominated eras. Furthermore, for the present model, it is demonstrated that the eventual fate of the Universe is mostly insensitive to the initial conditions, in contrast to the standard ΛCDM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Lambda \hbox {CDM}$$\end{document} model. In particular, the solution displays the properties of a dynamic attractor, which is reminiscent of quintessence and k-essence models. Subsequently, the cosmic coincidence problem is averted. The amount of deviation from a conserved stress energy–momentum tensor is shown to be more remarkable during the period when the dark energy evolves more rapidly. On the other hand, the conservation law is largely restored for the infinite past and future. The implications of the present approach are addressed.
Introduction
As an alternative to general relativity, Rastall gravity is characterized by the modified conservation law of the stress energy-momentum tensor (SET) in curved spacetime [1]. The theory implies intriguing novelty in various aspects regarding black hole physics [2][3][4][5][6][7][8][9][10][11][12] and cosmology [13][14][15][16][17][18][19][20][21][22] as it has been explored recently by many authors. In particular, the rudimentary feature of Rastall gravity, in a natural manner, supplies an alternative implementation for the dark energy. a e-mail<EMAIL_ADDRESS>(corresponding author) b e-mail<EMAIL_ADDRESS>The potential limit of general relativity has been systematically investigated on the largest scale against various observational data, namely, the supernova, large scale structure, and the cosmic microwave background (CMB) measurements. Among others, one of the most significant findings is the apparent accelerating expansion of the Universe, and subsequently, the dark energy scenario has become the most accepted premise regarding a satisfactory account for the experimental data. Moreover, it is deduced that the Universe at the present day is mostly composed of dark energy and dark matter. Subsequently, the physical properties, as well as the cosmic evolution of dark energy, become an increasingly active area in cosmology [23][24][25][26][27], due to its immediate connection with our understanding of the fundamental nature of the Universe.
Although the standard Cold Dark Matter ( CDM) model supplies a reasonable account for the observed properties of the cosmos, it also confronts several challenges such as cosmic coincidence problem and fine tuning problem. In this regard, alternative approaches are primarily carried out by modifying Einstein's field equations, which can be further divided into two distinct categories. The first type of model focuses on the properties of the matter field, which gives rise to dynamical dark energy models. In the literature, efforts along this train of thought consist of quintessence [28], tachyon [29], k-essence [30], phantom [31], Chaplygin gas [32], holographic dark energy [33][34][35], agegraphic dark energy [36,37], among others. The second type of approach, on the other hand, is motivated by generalizing the geometry in Einstein's general relativity. Such attempts include f (R) [38], f (T ) [39], f (R, T ) theory [40], Brans-Dicke theory [41], Gauss-Bonnet theory [42], Lovelock [43], and Horava-Lifshitz theories [44][45][46][47].
In general relativity, the SET is minimally coupled to the geometry. Consider a matter field that possesses a classical continuous symmetry, and a conserved current is implied according to the Noether theorem. However, as an infinitesimal symmetry transformation is made local, the action is no longer invariant, but rather it gives rise to a contribution associated with the Noether current. The above spacetime dependent transformation is a well-known procedure of introducing a gauge field into the theory. Here, the metric is playing the role of the gauge field for a diffeomorphism invariance, and the latter is related to the translation symmetry of the original theory. Subsequently, the Hilbert energy-momentum tensor, defined by the variation of the action of the matter field with respect to the metric, is conserved. In this context, it has been argued that the Rastall gravity can be viewed such that the curvature-matter coupling is implemented by a non-minimal fashion [48]. Therefore, the theory might be classified into the second category of modified gravity.
From a physical viewpoint, both the gravitationally induced particle production [49][50][51] and quantum effects in curved spacetime [52] might be associated with the violations of the usual conservation law of the SET. This particularly meaningful as it is understood that the conservation of SET does not lead to particle production [53]. From the viewpoint of relativistic kinetic theory, there is one more apparent mechanism even if the particle number is conserved, namely, the kinetic diffusive process. As it was pointed out in Ref. [54], the SET of the matter field is not conserved, as the evolution of the matter field is governed by the Fokker-Planck equation. Moreover, it can be shown that the divergence of the SET equals to a conserved four-current. In the study of cosmology, the above physical scenarios are relevant and evidently lead to important implications. In particular, the non-conserved part of the SET might give rise to the dark energy which, subsequently, is responsible for the present accelerating expansion of the Universe [14][15][16][17]21,22]. In Ref. [22], the authors studied the accelerating expansion of the Universe by employing a generalized Rastall theory. In particular, a non-minimal coupling between the geometry and a pressureless matter field is shown to lead the transition from the matter-dominated era to the accelerating expansion. The cosmic evolution is also investigated for homogeneous and isotropic flat Friedmann-Lemaître-Robertson-Walker (FLRW) metric in Ref. [55]. The model is shown to be equivalent to the particle creation mechanism in Einstein gravity in the framework of non-equilibrium thermodynamics.
The present study involves such an attempt to construct a reasonable scheme for cosmic evolution in a generalized Rastall gravity. In our model, the dark energy is implemented so that it is closely related to the violation of SET. The amount of violation is found to be more significant during the period when the contribution of dark energy increases and raises to its present value. It eventually becomes insignificant, as it is naturally dictated by the equations of motion. The resultant cosmic evolution experiences three stages, namely, radiation dominated, ordinary matter dominated, as well as dark energy and dark matter dominated eras. We also show that the eventual fate of the Universe is insensitive to the initial conditions, owing to the dynamical attractor behavior of the solution.
The rest of the paper is organized as follows. In the following section, we briefly discuss the generalized Rastall gravity utilized in the present study. The equations of motion of the cosmic evolution are derived in Sect. 3. Numerical results are presented in Sect. 4. Concluding remarks are given in the last section.
Generalized Rastall gravity
In Refs. [11,12], based on the original idea by Rastall [1], we proposed a generalized formulation of the Rastall theory. To be specific, the equation of the gravitational field equation and that of the SET read where κ = 8π G. We also impose a physical requirement that the effect of A μ ν and its derivatives must vanish in flat spacetime. In fact, it can be shown that the above formulation is rather general so that several modified gravity theories could be viewed as its special cases [11].
As for the purpose of the present study, we consider a specific case, namely, where H vanishes when R = 0. On the other hand, as a scalar, H can be a function of the Ricci scalar R, T ≡ g μν T μν and other constants. By substituting the form of A μν into Eq. (2.1), we have For algebraic convenience, one defines 3) can be rewritten in essentially the same form as in general relativity Although it is mathematically similar, usually, it is not physically appropriate to interpret τ μν as the SET of the matter field [56]. If one contracts both sides of the gravitational field equation, it gives Owing to the reasons to be discussed below, we choose where is to be determined shortly. We note that, in the vacuum, both factors on the numerator vanish as R → 0. In order that H is a well-defined quantity, one requires that the denominator of Eq. (2.7) being regular even when R → 0.
By substituting H into Eq. (2.6), one finds a quadratic algebraic equation, which implies the following two solutions for R: The first solution is not physically relevant, because here, we will investigate the scenario where R remains finite even when the matter field T μν vanishes. This is precisely the case where dark energy plays a significant role in cosmic evolution. Therefore, we will only explore the implication of the second solution. For the present model, this implies that is nonvanishing, while T μν vanishes. This, in turn, ensures that the denominator of Eq. (2.7) will be regular in our approach. By substituting it back into field equations, one finds Before proceeding further, we pause to give a few comments regarding Eq. (2.9). First, if one assumes ≡ eff /κ where eff is a constant, the above equations become identical to those of the standard CDM model. Therefore, it seems rather appealing to identify the physical content of with the cosmological constant. Although, in the present model, as further discussed below, its temporal dependence plays an essential role. In Ref. [12], it is demonstrated that an (anti-)de Sitter solution can be effectively found in Rastall gravity where the spacetime is asymptotically flat. It is achieved by taking H = H (R) and T μν = 0. In other words, the above solution again confirms the previous findings that a metric in asymptotically flat Rastall gravity naturally gives rise to that in general relativity with a cosmological constant. Moreover, according to the second equation of Eq. (2.9), measures the violation of the SET. Indeed, from the viewpoint of the Rastall gravity, all different types of matter fields are described by T μν , as a result, the observation of dark energy merely reflects, to what degree, the SET of the matter field deviates from a conserved current. It is also worth mentioning that Eq. (2.9) is very similar to those obtained from different theories where the conservation of the SET is partly breaking (for instance, see Refs. [57,58] and related discussions in the last section).
In the following section, we proceed to derive the equations for cosmic evolution and investigate their solutions.
Accordingly, we will treat as a variable, and solve its temperoal dependence.
Cosmic evolution in generalized Rastall gravity
The equations for cosmic expansion can be formulated by employing the co-moving coordinates, in terms of which the SET of the matter field is given by According to the discussions in the previous section, we denote ≡ ρ de , the energy density of the dark energy. It is noted, by using Eq. (2.4) and the solution Eq. (2.8), it is straightforward to show that the tensor τ μ ν reads where P de = −ρ de is recognized as the pressure of dark energy. In other words, although is not a constant, the equation of state of the dark energy still satisfies a simple form, namely, w de = P de ρ de = −1, which is in agreement with the observed results. Furthermore, the the cosmological principle implies that ρ = ρ(t), P = P(t), ρ de = ρ de (t), and P de = P de (t) are functions independent on spatial coordinates.
We proceed to derive the equations of motion in terms of the FLRW metric where k represents the curvature density of the Universe. Therefore, the (0, 0) and (1, 1) components of gravitational field equation in Eq. (2.9) can be rewritten as while the equation regarding the SET giveṡ We note only two of the above three equations are independent.
We consider the matter content of the Universe consists of radiation, ordinary matter, dark matter, and dark energy.
Radiation, ordinary matter, and dark matter are assumed to be independent between one another. They satisfy the standard equations of states, namely, P r = 1 3 ρ r and P m = P dm = 0. Therefore, the total pressure and density of the matter fields are given by As independent fluid components, we further assume that radiation and ordinary matter satisfy, respectively, an equation regarding the conservation of its SET, namely, For the dark matter, however, the corresponding equation is constrained by Eq. (3.5). It is not difficult to show that the resultant equation readṡ Now, there is only one free variable left, and for the last equation, we impose a rather simple scenario: which can be viewed as to effectively incorporate a specific type of interaction between the dark energy and dark matter. We note that this is in tune with the fact that Eq. (3.10) Here, the radiation and ordinary matter evolve as in standard CDM model. Also, the evolution of the dark energy accompanies that of dark matter, which reads Here, the index 0 indicates the values at present. One can also rewrite the field equation similar to the Friedman equation. By introducing the Hubble parameter H ≡ȧ a and and the spatial curvature density (3.12) one finds r + m + dm + de + k = 1, (3.13) where the i = 8π Gρ i 3H 2 with i = r, m, dm, de, k indicating the density parameters of radiation, ordinary matter, dark matter, dark energy, and spatial curvature respectively.
The deceleration parameter q ≡ −ä ȧ a 2 is found to be (3.14)
Numerical results
In the section, we present the numerical results in Figs. 1, 2, 3, 4, 5, 6, 7 and 8. We first determine the constants of the integration regarding equations of the cosmic evolution by the values of the measurements to date [59]. To be specific, we choose dm0 = 0.27, de0 = 0.68, m0 = 0.05. Also, we assume a spatially flat Universe by considering k = 0. Moreover, the redshift z = 1100, where the energy density of ordinary matter exceeds that of the radiation, is also taken as an input [60]. Subsequently, for the proposed model, the parameter β is found to be 2.52, which will be used in the remainder of this paper. The calculations are then carried out for the generalized Rastall theory, which are compared against those from the standard CDM model. The corresponding results obtained by adopting the above parameters for five variables a, ρ de , ρ dm , ρ r , and ρ m for given β. Subse- Fig. 4 The calculated deceleration parameter q as a function of a/a 0 . The present day a/a 0 = 1 is indicated by a vertical black solid line. The calculations are carried out for different parameters in generalized Rastall gravity. The cosmic evolution evaluated by using the specific initial conditions which reproduces the measurements is presented by solid curves. Those obtained by using different perturbed initial conditions are indicated by dashed and dotted curves Fig. 5 The same as Fig. 4. The calculated deceleration parameter q as a function of a/a 0 . The calculations are carried out for the standard CDM model. The cosmic evolution, as well as the results regarding arbitrary initial perturbations, are shown in solid, dashed, and dotted curves. The zoomed-in plot illustrates the deceleration parameters in the vicinity of a/a 0 = 1 quently, we investigate how the evolution of the composition of the Universe, and in particular, the density parameters at present day a/a 0 = 1, depends on different initial conditions. The latter are presented in dashed and dotted curves in Figs. 1, 2, 4, and 5 for both models.
As expected, from Fig. 1, the results show that the cosmic evolution consists of three stages, namely, the radiation dominated, ordinary matter dominated, as well as dark The dynamic attractor solution in the generalized Rastall gravity. The results show the deviations from the specific solution governed by a specific choice of initial conditions. The calculations are carried out for the differences in the density parameter of dark energy de (x-axis) and those in deceleration parameter q (y-axis). Each individual curve is obtained by evaluating the cosmic evolution with an arbitrary initial condition. The red dashed curve and black dotted curve correspond to the same perturbations investigated in Figs. 1 and 4. The calculations are carried out for generalized Rastall gravity by using the parameters given in the text energy and dark matter dominated eras. Also, it can be clearly inferred that the eventual fate of the Universe, calculated by the present model, is insensitive to the initial conditions. To be specific, the density parameters for the dark energy and dark matter all converge to the given values, irrelevant to specific initial conditions. Meanwhile, during the evolution, the compositions of the radiation and ordinary matter reflet the Fig. 8 The amount of deviation from a conserved SET, shown as a function of a/a 0 . The calculations are carried out for generalized Rastall gravity by using the model parameters described in the text details of the perturbed initial conditions. This point becomes particularly evident as one compares the above results against those of the standard CDM model shown in Fig. 2. In the DCM model, the density parameters at present a/a 0 = 1 are dictated largely by the initial conditions, as shown by the zoomed-in plot of Fig. 2. We note that the present findings are in agreement with other approaches [35,61], which incorporate the interaction between the dark energy and dark matter. The difference for the present model is that, in the framework of Rastall theory, the dark energy degree of freedom appears naturally from the deviation from the conservation law of the SET.
To clearly illustrate the difference in the resultant cosmic evolution between the two models, we present a comparison of the calculated density parameters in Fig. 3. It is found that although the density parameters of the dark energy and dark matter are identical at the present day in both models, their respective rates of change are distinct. In the CDM model, the density parameter increases rapidly at a/a 0 = 1, whereas that of the matter falls dramatically. As a result, to reproduce their measured values at the present day, one must carefully tune the initial conditions, which, in turn, gives rise to the related coincidence problem, as illustrated in Fig. 2. In the generalized Rastall theory, on the other hand, the evolutions of ordinary matter and dark matter are separated. The dark matter starts to arise together with the dark energy, owing to their interaction, after the ordinary matter dominated era. Moreover, both the dark energy and dark matter begin to saturate at the present day. Therefore their values do not sensitively depend on the initial conditions.
In Figs. 4, 5 and 6, one shows the resultant deceleration parameters for different initial conditions as functions of redshift in both models. Again, it is found that the deceleration parameter eventually approaches a given value, independent of specific initial conditions. Regarding both models, the values of q are identical at a/a 0 = 0, and the general trend is also found to be similar. However, for the Rastall gravity, one observes that q has begun to converge at a/a 0 = 0. This is different from the case of the CDM model where, again, at the present-day q is falling rapidly. As a result, the related value of q is sensitively governed by the specific initial conditions.
The above properties regarding the generalized Rastall theory can be shown more transparently as one focuses on the deviations from the specific solution discussed at the beginning of the section. The corresponding results are presented in Fig. 7 where one studies the discrepancies in cosmic evolutions by arbitrarily perturbing the initial conditions. To illustrate, we have chosen to show the differences in the density parameter of dark energy de and the deceleration parameter q. It is observed that the solution displays the properties of a dynamic attractor, which is reminiscent of quintessence and k-essence models. In other words, it is found that the deviations in evolution regarding different initial conditions all converge to the origin. Therefore, they are insensibility to the initial conditions in the present approach.
Last but not least, in Fig. 8, we show the amount of deviation from a conserved SET, which is the 0-component of the r.h.s. of Eq. (2.1), as a function of a/a 0 , for the generalized Rastall theory. As discussed above, for the present model, the amount of violation is related to the dynamical evolution of the dark energy. As shown in Fig. 8, the deviation is timedependent. Its magnitude becomes more significant when the dark energy evolves more rapidly, and the peak is found to locate at approximately a/a 0 ∼ 0.3. On the other hand, the SET is mostly conserved in the infinite past and future.
Discussions and concluding remarks
Owing to the fact that one has to discard one of the solutions of Eq. (2.8), which introduces a vanishing factor on both sides of Eq. (2.7), the numerator of the equation is chosen as a second-order polynomial. In fact, Eq. (2.7) only contains as an unknown scalar function, which is identified with a dynamical cosmological constant. For this reason, it is actually a rather economical choice of ansatz in the present model.
In comparison to the standard CDM model, effectively, the proposed scheme only contains one additional variable, . The latter is described by the assumed equation of motion Eq. (3.9). In this context, it is a minimal scheme necessarily to describe the dynamical evolution of dark energy. In comparison to other recent studies [22,55] about cosmic evolution in Rastall theory, the present approach introduces a unified scheme to deal with different matter contents of the Universe.
In other words, by solving a closed system of equations, different eras of cosmic evolution are derived naturally. Moreover, we argue that our model possesses a dynamic attractor solution, which provides a possible explanation for the coincidence problem.
It is also interesting to mention that the non-conserved SET can be treated in terms of a generalized version of the two measure theories [62,63]. In this case, the dynamics can be derived from an action which consists of two measures. In particular, the latter involves a scalar density in the place of the usual factor of the Jacobian √ −g. The theory is recently generalized in order to accommodate the fact the SET is not conserved as one considers the diffusive process in the relativistic Fokker-Plank equation [54]. There, the divergence of the SET is shown to be related to the conserved particle flow. This can be achieved by replacing the dynamic spacetime four-vector in the original theory by the gradient of a scalar field [57,58]. The resultant theory gives rise to a unified description of the interacting dark energy and dark matter. It is, therefore, intriguing to compare the above approach against the Lagrangian formalism of Rastall theory.
To summarize, the present study involves an attempt to propose a scheme for cosmic evolution in a generalized Rastall gravity. In our model, the physical content of the dark energy is attributed to the non-conserved sector of the SET. The resultant cosmic evolution is naturally found to consists of three stages, namely, radiation dominated, ordinary matter dominated, as well as dark energy and dark matter dominated eras. Also, for the present model, it is shown that the eventual fate of the Universe is largely insensitive to the initial conditions, and the cosmic coincidence problem is therefore averted. Furthermore, we show that the amount of violation is found to be more significant when the dark energy evolves dynamically.
Data Availability Statement
This manuscript has no associated data or the data will not be deposited. [Authors' comment: The data regarding the parameters of the cosmic evolution are taken from the respective references. The numerical results presented in the paper can be readily obtained by solving the system of equations Eqs. (3.10)-(3.14) given in the paper with the initial conditions described therein.] Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indi-cated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecomm ons.org/licenses/by/4.0/. Funded by SCOAP 3 . | 5,853.4 | 2020-06-01T00:00:00.000 | [
"Physics"
] |
Phage Cocktail in Combination with Kasugamycin as a Potential Treatment for Fire Blight Caused by Erwinia amylovora
Recently, there has been an increasing number of blight disease reports associated with Erwinia amylovora and Erwinia pyrifoliae in South Korea. Current management protocols that have been conducted with antibiotics have faced resistance problems and the outbreak has not decreased. Because of this concern, the present study aimed to provide an alternative method to control the invasive fire blight outbreak in the nation using bacteriophages (phages) in combination with an antibiotic agent (kasugamycin). Among 54 phage isolates, we selected five phages, pEa_SNUABM_27, 31, 32, 47, and 48, based on their bacteriolytic efficacy. Although only phage pEa_SNUABM_27 showed host specificity for E. amylovora, all five phages presented complementary lytic potential that improved the host infectivity coverage of each phage All the phages in the cocktail solution could lyse phage-resistant strains. These strains had a decreased tolerance to the antibiotic kasugamycin, and a synergistic effect of phages and antibiotics was demonstrated both in vitro and on immature wound-infected apples. It is noteworthy that the antibacterial effect of the phage cocktail or phage cocktail-sub-minimal inhibitory concentration (MIC) of kasugamycin was significantly higher than the kasugamycin at the MIC. The selected phages were experimentally stable under environmental factors such as thermal or pH stress. Genomic analysis revealed these are novel Erwinia-infecting phages, and did not encode antibiotic-, virulence-, or lysogenic phage-related genes. In conclusion, we suggest the potential of the phage cocktail and kasugamycin combination as an effective strategy that would minimize the use of antibiotics, which are being excessively used in order to control fire blight pathogens.
Introduction
The bacterium Erwinia amylovora is a causative agent of fire blight, a devastating disease of rosaceous plants [1,2]. Fire blight-free regions suffer devastating economic losses following the first outbreak of fire blight invasion due to there being no specific methods to effectively control plant pathogens, except for a limited number of antibiotics such as streptomycin, oxytetracycline, and kasugamycin [3][4][5]. E. amylovora isolates from apple orchards are known to have resistance to streptomycin, the primary treatment for fire blight [6]. Furthermore, the high prevalence of resistance genes to these antibiotics in the environment (endosphere, rhizosphere, or phyllosphere), creates a high probability of the transfer of antibiotic resistance genes to pathogens [7][8][9]. Consequently, a high concentration of antibiotics should be used to be effective against bacterial outbreaks, including fire blight, which can cause dysbiosis in the environmental microbiota. A disrupted microbial balance can facilitate an outbreak of diseases [10,11].
To overcome the problem of antibiotic resistance in E. amylovora, a number of alternatives have been reported, such as essential oil, plant extracts, and antagonistic bacteria [12][13][14][15][16][17]. In addition, bacteriophages (phages) have been suggested as potential alternatives to antibiotics for controlling fire blight owing to their direct killing effect [18,19]. The comparative advantages of using phages to control pathogens mainly comprise their ability to specifically recognize cell surface receptors on their bacterial hosts to infect and lyse the pathogen after replication within the host cell [20,21]. For decades, a number of phages have been characterized as effective agents against fire blight, and several commercial phages have been developed and made available worldwide as solutions against fire blight such as Omnilyticus AgriPhage™-Fire Blight (Salt Lake City, UT, USA) and Enviroinvest Erwiphage PLUS (Kertváros, Hungary) [22][23][24]. The high specificity of therapeutic phages confers on them the advantage of being able to be used as a biocontrol method without affecting beneficial microbes in the environment. However, it can be a major limitation at the same time owing to the inability of such phages to act on a broad range of pathogens. Therefore, phages with a broad host range are generally preferred for therapeutic use [25][26][27].
Combining different phages in cocktail solutions is the primary strategy to overcome the limitation of the narrow host range of phage therapy [28,29]. Different phages in cocktail solutions can complement the host range coverage of every other phage in the solution, as well as address the issue of phage resistance being developed due to the administration of a single type of phage [30,31]. In particular, phage cocktails are expected to show a synergistic effect of the combination of phages, although this is not always observed to be the case [32,33]. Another strategy is to combine antibiotics with phages [34]. Synergy between phages and antibiotics can be demonstrated to occur by the observation of enhanced plaque size or clarity, and improved growth characteristics of phages, such as a shortened eclipse period or increased burst size [35][36][37][38][39]. The application of phages to control bacterial pathogens can therefore reduce the excessive use of antibiotics, thus, allowing them to be reserved for urgent clinical needs.
The present study investigated the biocontrol potential of newly isolated Erwinia phages. With the five phages selected in our study, we showed the effectiveness of the resultant phage cocktail, as well as that of its combination with antibiotics, which we propose as an alternative strategy to control fire blight caused by E. amylovora.
Morphological and Biological Characteristics of the Bacteriophages
The selected phages were morphologically recognized as belonging to the family Myoviridae ( Figure 1). Extended long tail fibers were observed around ϕ48 (Figure 1e). Structural observations of phages ϕ27, ϕ31, ϕ32, ϕ47, and ϕ48, showed the presence of a capsid having diameter minimum 68 Figure S1).
The selected phages were morphologically recognized as belonging to the family Myoviridae ( Figure 1). Extended long tail fibers were observed around φ48 (Figure 1e). Structural observations of phages φ27, φ31, φ32, φ47, and φ48, showed the presence of a capsid having diameter minimum 68.5 ± 2.76 nm and maximum 139.15 ± 5.47 nm, and a contractile tail having length minimum 115.1 ± 2.16 nm and maximum 196.32 ± 11.45 nm (n = 5) ( Table 1). The host range of the five selected phages is represented in using 94 and 25 isolates of E. amylovora and E. pyrifoliae, respectively. All five phages infected 100% of the E. amylovora strains (94/94) recently isolated in South Korea. Although φ27 showed a narrow host range when tested against E. pyrifoliae strains (8/25; 32%), other phages could complement the host coverage, rendering all E. pyrifoliae strains susceptible to the infectivity of those phages (Supplementary Materials Figure S1).
In Vitro Bacterial Killing Assay
Phage administration led to a rapid lysis of E. amylovora ( Figure 2). Each phage was effective in lysing E. amylovora up to 8 h; however, the regrowth of E. amylovora was observed at 24 h of incubation with φ27, φ47, and φ48. The CFU of regrown bacteria in samples treated with the cocktail, φ31, and φ32 was significantly lower than that of the samples treated φ27, φ47, and φ48 (p < 0.05). The phage cocktail contained 1/5 parts concentration of each phage, and yet was an extremely effective solution for inhibiting the pathogen. The administration of the single phages resulted in 2.4, 3.5, 3.5, 1.2, 1.4 log CFU/mL reduction in the final bacterial counts of phages φ27, 31, 32, 47, and 48, respectively, and
In Vitro Bacterial Killing Assay
Phage administration led to a rapid lysis of E. amylovora ( Figure 2). Each phage was effective in lysing E. amylovora up to 8 h; however, the regrowth of E. amylovora was observed at 24 h of incubation with ϕ27, ϕ47, and ϕ48. The CFU of regrown bacteria in samples treated with the cocktail, ϕ31, and ϕ32 was significantly lower than that of the samples treated ϕ27, ϕ47, and ϕ48 (p < 0.05). The phage cocktail contained 1/5 parts concentration of each phage, and yet was an extremely effective solution for inhibiting the pathogen. The administration of the single phages resulted in 2.4, 3.5, 3.5, 1.2, 1.4 log CFU/mL reduction in the final bacterial counts of phages ϕ27, 31, 32, 47, and 48, respectively, and the bactericidal effect of the five-phage cocktail led to a 3.7 log CFU/mL reduction of the bacterial count, which is a significant decrease compared with the bacterial count of the untreated control group (p < 0.001 at 2 h, 4 h, 6 h, 8 h, and 24 h).
Biological Characteristics of Phage-Resistant Erwinia amylovora TS3128 Derivatives
The profile of phage susceptibility of single phage-resistant strains is summarized in Figure 3a. The R27 strain was susceptible to phages ϕ31, 32, 47, and 48. While the phage resistance of R31, R32, R47, and R48 was induced by ϕ31, 32, 47, and 48, respectively, the resistant strains gained cross-resistance to all the other unrelated phages (ϕ31, 32, 47, and 48) except ϕ27 (Figure 3a). However, the cocktail solution infected all phage-resistant strains. the bactericidal effect of the five-phage cocktail led to a 3.7 log CFU/mL reduction of the bacterial count, which is a significant decrease compared with the bacterial count of the untreated control group (p < 0.001 at 2 h, 4 h, 6 h, 8 h, and 24 h).
Figure 2.
In vitro bactericidal effect of Erwinia phages φ27, φ31, φ32, φ47, φ48, and their cocktail. The viable bacterial cells were counted over 24 h. The E. amylovora strain TS3128, a reference strain for research in Korea, was used. The bars of each point indicate the standard deviation. Statistical significance was calculated using the one-way analysis of variance test with Tukey post-hoc, and the significance threshold was set at p < 0.05. Means at the same sampling time point with different letters (a-e) are significantly different.
The minimal inhibitory concentration (MIC) values of kasugamycin against the wild type E. amylovora TS3128 and the phage-resistant strains are shown in Figure 3b. The MIC of kasugamycin for wild type (WT) TS3128 was observed to be 64 μg/mL. Moreover, we observed a 2-to 4-fold decrease in the MIC of the phage-resistant strains R27, R31, and R32.
Figure 2.
In vitro bactericidal effect of Erwinia phages ϕ27, ϕ31, ϕ32, ϕ47, ϕ48, and their cocktail. The viable bacterial cells were counted over 24 h. The E. amylovora strain TS3128, a reference strain for research in Korea, was used. The bars of each point indicate the standard deviation. Statistical significance was calculated using the one-way analysis of variance test with Tukey post-hoc, and the significance threshold was set at p < 0.05. Means at the same sampling time point with different letters (a-e) are significantly different.
Phage-Antibiotic Synergy (PAS) Assay
To determine the enhanced antibacterial activity of the phage cocktail with kasugamycin, a phage-antibiotic synergy assay was performed with different antibiotic concentrations and the bacteria TS3128 ( Figure 4). Kasugamycin at MIC inhibited the growth of The minimal inhibitory concentration (MIC) values of kasugamycin against the wild type E. amylovora TS3128 and the phage-resistant strains are shown in Figure 3b. The MIC of kasugamycin for wild type (WT) TS3128 was observed to be 64 µg/mL. Moreover, we observed a 2-to 4-fold decrease in the MIC of the phage-resistant strains R27, R31, and R32.
Phage-Antibiotic Synergy (PAS) Assay
To determine the enhanced antibacterial activity of the phage cocktail with kasugamycin, a phage-antibiotic synergy assay was performed with different antibiotic concentrations and the bacteria TS3128 ( Figure 4). Kasugamycin at MIC inhibited the growth of E. amylovora, while sub-MIC inoculations allowed bacterial growth. There was a slight enhancement in the antibacterial effect from using the phage cocktail and 1/4 MIC kasugamycin combination. The advanced effect was much higher when the phage cocktail was combined with 1/2 MIC and 1 MIC kasugamycin. The final viable bacterial cell count reduction was 3.7 (phage cocktail), 3.8 (phage cocktail-1/4 MIC kasugamycin), 5.1 (phage cocktail-1/2 MIC kasugamycin), and 5.4 (phage cocktail-1 MIC kasugamycin), therefore resulting in the PAS effect (difference of bacterial cell count reduction between phage cocktail only and phage cocktail-kasugamycin combination) to be 0.1 (phage cocktail-1/4 MIC kasugamycin), 1.4 (phage cocktail-1/2 MIC kasugamycin), and 1.7 (phage cocktail-1 MIC kasugamycin). The samples treated using the phage cocktail with 1/4 MIC, 1/2 MIC, and 1 MIC kasugamycin showed a significant reduction compared with treated with kasugamycin alone (p < 0.001).
Experiment on Apple Fruit under Controlled Conditions
A biocontrol assay of the phages was conducted for the phage cocktail and its combination with antibiotics ( Figure 5
Experiment on Apple Fruit under Controlled Conditions
A biocontrol assay of the phages was conducted for the phage cocktail and its combination with antibiotics ( Figure 5). A significant improvement in inhibition of bacterial growth was observed when the phage cocktail-kasugamycin combination with 1/2 MIC and 1 MIC was administered compared with the phage cocktail treatment at day 4 and 6 (p < 0.001).
Stability Assay
The stability of the phages under environmental stressors (pH and temper examined. A majority of the five phages were considerably stable under the th ditions tested (4-50 °C), except for phages φ32 and φ48, which are both slightly to high temperatures (50 °C; Figure S2). In addition, the infectivity of φ32 w hindered under alkaline conditions, while the other phages were stable unde pH conditions ranging from pH 4 to 9 ( Figure S2).
Genomic Analysis of the Selected Phages
The general features of the genomes of the phages are presented in Table Erwinia phages φ27, φ31, φ32, φ47, and φ48 possessed double-stranded circ having a guanine-cytosine (GC) content of 44.07%, 49.53%, 49.19%, 34.48%, a respectively. Phage φ27 possessed a relatively small genome (53,014 bp) com
Stability Assay
The stability of the phages under environmental stressors (pH and temperature) was examined. A majority of the five phages were considerably stable under the thermal conditions tested (4-50 • C), except for phages ϕ32 and ϕ48, which are both slightly vulnerable to high temperatures (50 • C; Figure S2). In addition, the infectivity of ϕ32 was slightly hindered under alkaline conditions, while the other phages were stable under different pH conditions ranging from pH 4 to 9 ( Figure S2).
Genomic Analysis of the Selected Phages
The general features of the genomes of the phages are presented in Table 1. The five Erwinia phages ϕ27, ϕ31, ϕ32, ϕ47, and ϕ48 possessed double-stranded circular DNA, having a guanine-cytosine (GC) content of 44.07%, 49.53%, 49.19%, 34.48%, and 49.52%, respectively. Phage ϕ27 possessed a relatively small genome (53,014 bp) compared with the those of other phages, and the Erwinia jumbo phage ϕ47 had a large genome (355,376 bp). In total, 78, 337, 336, 540, and 358 open reading frames (ORFs) were identified in the genomes of ϕ27, 31, 32, 47, and 48, respectively. The genus of ϕ27 was identified as Loessnervirus, characterized by a genome of 55.80 kbp with 44.2% GC content, such as the Erwinia phage vB_EamM-Y2 and the Pantoea phage vB_PagM_SSEM1. No encoded tRNAs have been previously reported in the genomes of Loessnervirus; however, one tRNA was identified in the genome of ϕ27. The genus of ϕ31 and ϕ32 was identified as Alexandravirus, represented by the Erwinia phage Alexandra and the Dickeya phage AD1. This genus presents genomes of 261-266 kbp coding two distinct tail sheath proteins. Phages ϕ31 and ϕ32 have two tail sheath proteins and no tRNA. The genome of ϕ47 was identified as Eneladusvirus, represented by the Serratia phage BF and the Yersinia phage Yen9-04. This genus presents a genome of 354-357 kb with 34.4% GC contents and 35 tRNAs. Phage ϕ48 has two tRNAs but its genome was unclassified.
The Erwinia phages in this study showed dissimilar and unique genomic arrangements, except for phages ϕ31 and ϕ32, as they were in the same genus. Even though most of the predicted ORFs had no matches in any database, identified proteins from the five phages could be categorized into the following six groups based on their functions: proteins related to structure and packaging, nucleotide metabolism, tRNA, lysis, additional functions, and hypothetical proteins ( Figure 6, Tables S2-S6). The Erwinia phages in this study showed dissimilar and unique genomic arrangements, except for phages φ31 and φ32, as they were in the same genus. Even though most of the predicted ORFs had no matches in any database, identified proteins from the five phages could be categorized into the following six groups based on their functions: proteins related to structure and packaging, nucleotide metabolism, tRNA, lysis, additional functions, and hypothetical proteins ( Figure 6, Tables S2-S6). The open reading frames were functionally assorted into six groups of proteins related to: structure and packaging (blue), nucleotide metabolism (yellow), lysis (red), and additional functions (purple), as well as tRNA proteins or tRNA-related proteins (black), and hypothetical proteins (gray). Scale is base pair (bp).
Comparative Genomic Analysis
The whole-genome sequences of the five phages were evaluated for comparative analysis with representative phages infecting Erwinia spp., Dickeya spp., Pantoea spp., and Pectobacterium spp. A phylogenetic analysis using the Virus Classification and Tree Building Online Resource (VICTOR) clustered the phages according to their taxonomy (Figure 7a). Phage ϕ27 was clustered with Erwinia phage vB_EamM-Y2 (NC 019504.1) and Pantoea phage vB_PagM_SSEM1 (NC 048875.1), in a manner similar to the clustering exhibited by Loessnervirus. The cluster comprising ϕ31 and ϕ32 was clustered with Dickeya phage vB_DsoM_AD1 (NC 048054.1), and these two phages were identified as Alexandravirus.
Phage ϕ47 was clustered with Pectobacterium phage CBB (NC_041878.1) and identified as Eneladusvirus. Phage ϕ48 formed a distinct cluster that diverged from a common ancestor with Agricanvirus bacteriophages. The dot plot analysis of the 79 phages indicated firm clustering and supported the phylogenetic analysis (Figure 7b). Phage ϕ27 had a strong lineage association with Loessnervirus (Erwinia phage vB_EamM-Y2 and Pantoea phage vB_PagM_SSEM1); phages ϕ31 and ϕ32 were seen to be closely related to Alexandravirus (Erwinia phage vB_EamM_Alexandra and Dickeya phage vB_DsoM_AD1). In contrast, phages ϕ47 and ϕ48 did not demonstrate close relatedness with other reported Erwinia phages.
Progressive Mauve was used to align and compare phages ϕ27, ϕ31, ϕ32, ϕ47, and ϕ48 with genetically close phages: Pantoea phage vB_PagM_SSEM1, Dickeya phage vB_DsoM_AD1, Erwinia phage vB_EamM_Alexandra, Pectobacterium phage CBB, and Erwinia phage vB_EamM_RAY (Figure 7c). The genome of ϕ27 and Pantoea phage vB_PagM_ SSEM1 were identified as the Loessnervirus genus. The genomes of ϕ31, Dickeya phage vB_DsoM_AD1, ϕ32, and Erwinia phage vB_EamM_Alexandra were closely related with the genus Alexandravirus. Furthermore, close relatedness of ϕ47 with Pectobacterium phage CBB was also determined. A comparative study between the genomes of ϕ48 and Erwinia phage vB_EamM_RAY (Agricanvirus) was conducted, since the genus of ϕ48 was not identified in the genomic analysis; this showed similarity with, however, relevant differences. The results showed that the genome sequences of ϕ27, ϕ31, ϕ32, ϕ47, and ϕ48 presented the differences from their closest relatives, which supported the comparative results from the phylogenetic analysis and dot plot analysis.
Discussion
Fire blight was first reported in 2015, and since then there has been an increasing number of outbreaks in South Korea, especially recently [40,41]. Without any regulations regarding the administration order for antibiotics to control the fire blight in South Korea, secondary agents, including kasugamycin, are widely used in general. There are no investigations that reported the antibiotic resistance of E. amylovora in South Korea, however, misuse of the antibiotic agents can promote the evolution of resistance, and dysbiosis of the orchard environment, which would lead to the failure of the fire blight management of the nation. To combat this severe blight disease, our research team has been dedicated to developing phages as effective alternatives to antibiotics. Due to the presence of two nearly indistinguishable pathogens, E. amylovora and E. pyrifoliae, in South Korea, phages that are capable of infecting both pathogens are considered ideal biocontrol agents.
Although the phages used in this study were isolated using E. amylovora as their host, they could infect E. pyrifoliae, an endemic species that also led to blight symptoms in plants in South Korea, which is in accordance with the previous reports that Erwinia amylovora bacteriophages have a broad host range [22,23,[42][43][44][45][46]. From our Erwinia phage isolates, we screened phages based on their bacterial cell lysis efficacy and selected phages φ27, φ31, φ32, φ47, and φ48 to form the Erwinia phage cocktail solution. Phages in the cocktail improved each other's host range complementarily, leading the cocktail to be infective towards all recently isolated E. amylovora and E. pyrifoliae strains. Combining phages with complementary host ranges is one of the key virtues of phage cocktails, since phages present host-specific infectivity [28].
Discussion
Fire blight was first reported in 2015, and since then there has been an increasing number of outbreaks in South Korea, especially recently [40,41]. Without any regulations regarding the administration order for antibiotics to control the fire blight in South Korea, secondary agents, including kasugamycin, are widely used in general. There are no investigations that reported the antibiotic resistance of E. amylovora in South Korea, however, misuse of the antibiotic agents can promote the evolution of resistance, and dysbiosis of the orchard environment, which would lead to the failure of the fire blight management of the nation. To combat this severe blight disease, our research team has been dedicated to developing phages as effective alternatives to antibiotics. Due to the presence of two nearly indistinguishable pathogens, E. amylovora and E. pyrifoliae, in South Korea, phages that are capable of infecting both pathogens are considered ideal biocontrol agents.
Although the phages used in this study were isolated using E. amylovora as their host, they could infect E. pyrifoliae, an endemic species that also led to blight symptoms in plants in South Korea, which is in accordance with the previous reports that Erwinia amylovora bacteriophages have a broad host range [22,23,[42][43][44][45][46]. From our Erwinia phage isolates, we screened phages based on their bacterial cell lysis efficacy and selected phages ϕ27, ϕ31, ϕ32, ϕ47, and ϕ48 to form the Erwinia phage cocktail solution. Phages in the cocktail improved each other's host range complementarily, leading the cocktail to be infective towards all recently isolated E. amylovora and E. pyrifoliae strains. Combining phages with complementary host ranges is one of the key virtues of phage cocktails, since phages present host-specific infectivity [28].
The ideal strategy for phage cocktails is to generate synergy between phages [28]. As the phages inhibit the secondary infection (superinfection) of their close relatives, it is a crucial factor to exclude the ones revealing the antagonistic effect in the cocktail [28]. One promising way to generate synergism is combining the phages having virion-associated enzymes [32]. In line with the prediction that the genome of pEa_SNUABM_47, a constituent of cocktail, encodes for tail spike lysozyme [45]. Indeed, pEa_SNUABM_47 revealed the synergistic effect in the first phases (0 to 8 h) of the in vitro bacterial killing assay with the phages that are genetically distant (Figures 2 and 6). Even though this effect could not be achieved over the long term (24 h), the selected phages did not show an antagonistic effect, which is not recommended for cocktail constituents [28].
Analysis of phage resistance in the five phages showed cross-resistance between ϕ31, ϕ32, ϕ47, and ϕ48 (Figure 3). Only the ϕ27-resistant strain (R27) did not show crossresistance with other phages, and vice versa. It is remarkable that phages selected from distinct genera could be cross-resistant (Figure 7a). The genomic arrangement of the five phages was totally unrelated; eventheir lysis-related proteins did not show homology to each other (Figure 7b,c, and Tables S2-S6). The infection process of the phages was considered to be the origin of the phage resistance and the cross-resistance. However, this contradicts previous presumptions of infection mechanisms differing based on the taxonomical status (family) of the phages [45]. As of today, a number of phages have been reported, and genomic classifications have been improved and updated. In our study, all phages were classified in the family Myoviridae; however, the host recognition strategy of the myophages with a small genome is presumed to differ from that of jumbo myophages. More detailed analyses are warranted in future studies to elucidate host-phage interactions. We suggest that analyzing cross-resistance patterns among candidate phages for cocktail solutions should be considered as the highest priority. Because the phages' host preference and infection have a dependency on exopolysaccharides (EPS) produced by E. amylovora, a novel strategy combining the strains that produce different amounts of EPS have been suggested for the host range analysis [46,47].
Even though resistance to antimicrobial agents is a major concern, the phages in our cocktail solution could control the phage-resistant strains (Figure 3a). Bacterial pathogens might acquire phage resistance by fitness trade-off [48]. To escape contact with phages, bacteria modify (or even lose) receptors used for phage infection as their first-line antiphage defense strategy [49]. Often, these alterations cause lowered viability, decreased pathogenicity, and increased susceptibility to antimicrobial agents [50]. Interestingly, a trade-off between phage resistance and kasugamycin susceptibility was observed in the phage-resistant E. amylovora strains R27, R31, and R32. The decreased MIC is indicative of PAS against E. amylovora. Indeed, the phage-antibiotic combination proved to have superior efficacy in both the in vitro and apple fruit assays, which may reduce the use of antibiotics in the field. PAS was observed even at sub-inhibitory antibiotic concentrations (Figures 4 and 5).
Aminoglycoside antibiotics, such as gentamicin, kanamycin, streptomycin, and kasugamycin, are translation-interfering drugs that can also hinder translation in phages, resulting in premature lysis [51,52]. Even worse from the perspective of phage therapy, in the long term aminoglycosides can cause the extinction of phages from the environment [53]. However, the antibiotic action of kasugamycin is competitive [54], therefore translation can be initiated if surplus initiation factors are present. Such translational initiators include initiation tRNA (tRNAi), such as tRNA-fMET, which is encoded in the jumbo Erwinia phage pEa_SNUABM_47 [45]. The synergy and facilitation between phages and kasugamycin is presumed to originate from the phage-originated translational initiator in the following process: (1) kasugamycin inhibits bacterial growth by interfering with translation; (2) phages infect stationary-phase bacteria and transcribe their genome, including tRNAi; (3) tRNAi of phage origin hijacks the translational machinery by competition and starts to translate phage proteins, allowing progeny release and propagation; and (4) phage replication continues while the adjacent bacterial cells are still in the stationary phase due to kasugamycin. Although the mechanism might not be exactly the same, a PAS effect has been hypothesized between gentamicin (another aminoglycoside antibiotic) and a Staphylococcus phage [55]. The tRNAs of jumbo phages increase phage fitness by improving the translational efficiency or independence of translation from the host factors [56]. Thus, the combination of phages encoding tRNAi and kasugamycin should be included as a biocontrol agent against fire blight.
Considering the findings of a previous report elucidating the importance of administration order in devising combined treatments with phages and antibiotics [57], the next step would be optimization of the administration order with the concentrations obtained in the present study (8 log plaque forming unit [PFU]/mL and sub-MIC of kasugamycin). We proposed the use of PAS for optimizing strategies to control E. amylovora and, consequently, fire blight and strategies involving PAS can reduce the excessive use of antibiotics in fire blight control. This can minimize the emergence and spread of antibiotic resistance among opportunistic pathogens present in the environment [58]. We propose our phage cocktail, and its combination with kasugamycin, to be an effective protocol to control the current blight outbreaks caused by Erwinia in South Korea, as the pathogens tested in our study are recently recovered strains from diseased plant tissue obtained from locations across South Korea. Further studies investigating the synergistic mechanisms of kasugamycin, and phages having their own translational initiator, are expected to broaden our options for alternative antibacterial strategies and reduce the excessive use of antibiotic agents.
Phage Isolation
A total of 220 samples were collected comprising 94 soil samples and 126 water samples from the area affected by the fire blight outbreak in South Korea, and the phages infecting E. amylovora were isolated from the samples using a protocol described in previous studies [59,60]. The Erwinia amylovora TS3128 strain, a reference strain for research in Korea, was cultured with exponential growth, and the samples were added to the cultures in a one-to-one ratio. The mixed samples were cultured at 27 • C for 24 h to amplify the phages. Samples presenting plaques were identified, collected, and subsequently filtered through a 0.45 µm syringe filter. The double-layer agar (DLA) method was used to confirm the bacteriolysis induced by the phages [61]. Cloning of the phages from the plaques was carried out five times to purify and isolate the respective phages.
Phage Propagation and Purification
The DLA method was used to amplify the phages, based on a protocol described in a previous publication [62]. The top agar layer was collected in an SM buffer (50 mM Tris [pH 7.5], 100 mM NaCl, and 10 mM MgSO 4 ) and mixed for 1 h. The mixture was centrifuged, and the supernatant was filtered through a 0.45 µm syringe filter to eliminate contaminants. Then, a polyethylene glycol/NaCl solution was added to the sample to precipitate the phage particles. The cesium chloride (CsCl) density gradient centrifugation method was used to purify the phage particles [45]. Phage samples with gradient layers of CsCl solution were ultracentrifuged for 3 h at 50,000× g using a Type 70 Ti fixedangle titanium rotor (Beckman, Brea, CA, USA). The sedimentation bands were collected and dialyzed using a 7000 MWCO Slide-A-Lyzer ® Dialysis Cassette (Thermo Scientific, Waltham, MA, USA). The purified samples (>10 10 PFU/mL) were stored at 4 • C for further analysis.
Transmission Electron Microscopy
The purified phage samples were attached for 1 min on separate glow-discharged TEM Grid FCF200-CU-50 Formvar/Carbon grids (Sigma-Aldrich, Burlington, MA, USA). After removing the sample solution, 2% phosphotungstic acid was added to the grids to stain the phages for 30 s, and the remaining solution was eliminated. The grids were air-dried for 1 h, and morphological study of the phages was performed using a Talos L120C transmission electron microscope (FEI, Hillsboro, OR, USA) operated at 120 kV. Three isolated virions were measured, and the mean size of the phages was calculated.
Bacteriophage Screening Assay
Bacteriophages were screened in two stages to select five effective phages based on their growth inhibition potential. Growth inhibition was determined based on optical density (OD) at 600 nm after 24 h of phage-bacteria co-culture. The initial screening was performed at 10 5 CFU/mL, and the second screening was performed with 10 6 CFU/mL of E. amylovora. The tests were performed in a 96-well plate with 10 8 PFU/mL of each phage and incubated at 27 • C with shaking (150 rpm). The phages and bacteria were prepared in nutrient broth. The growth inhibition was calculated as follows: % growth = OD600 of challenge OD600 of untreated host ×100 (1)
Bacteriophage Host Range Assay
A total of 94 strains of E. amylovora and 25 strains of E. pyrifoliae were tested to identify the host infectivity of the selected five phages: ϕ27, ϕ31, ϕ32, ϕ47, and ϕ48. The infectivity of phages was determined by performing a spot assay against recently recovered strains obtained from diseased plant tissue in South Korea. Serial dilutions of phage lysate (10 µL) at a concentration of 10 1 to 10 8 PFU/mL were added dropwise on the bacterial lawns, and the infectivity was represented as the efficiency of plating (EOP) value. The protocol was described in a previous study, with minor modifications, i.e., using a 52 • C water bath instead of a 46 • C heating block [63].
Bacterial Killing Assay In Vitro
The bactericidal efficacy of individual phages, and of their cocktail, was examined using E. amylovora TS3128 according to a method described in a previous publication, with minor modifications [64]. The strain (10 5 CFU/mL) was infected with phages at a concentration of 10 8 PFU/mL. The cocktail comprised identical ratios (1:1:1:1:1) of 2 × 10 7 PFU/mL of each phage. The mixtures were cultured at 27 • C with shaking (150 rpm), and the cell counts were observed over time. Each experiment was performed in triplicates (n = 3).
Phage Resistance Assay
The phage resistance assay was performed as previously described, with minor modifications [31]. After the in vitro bacterial killing assay, the surviving colonies were subcultured thrice to remove the residual phages. Then, phage susceptibility was tested as described above. If plaques were not observed, the strain was confirmed to be phage resistant. Phage-resistant strains were designated as follows; ϕ27-resistant strain (R27), ϕ31-resistant strain (R31), ϕ32-resistant strain (R32), ϕ47-resistant strain (R47), and ϕ48resistant strain (R48). The susceptibility of phages was determined using the five phages and the cocktail at a concentration of 2 × 10 9 PFU/mL. Ten microliters of serial dilutions (10 −1 to 10 −8 ) of phage solution were spotted on each phage-resistant strain: R27, R31, R32, R47, and R48. Negative control (N) and wild type (WT) were also tested.
Minimum Inhibitory Concentration (MIC) Assay
The MIC value of kasugamycin against the wild type E. amylovora and phage-resistant strains was determined using the broth microdilution method [65]. Serial dilutions (twofold) starting with 512 µg/mL were inoculated with the same volume of the bacterial solution (2 × 10 5 CFU/mL) and incubated for 24 h at 27 • C. The MIC of the antibiotics was determined by measuring the OD at 600 nm in triplicates (n = 3). The growth inhibition was calculated as follows and the results were visualized in a heatmap: % growth = OD600 of challenge OD600 of untreated host ×100 (2)
Phage-Antibiotic Synergy Assay
The advanced effect between the phage cocktail and kasugamycin was determined using E. amylovora with a method described in a previous study [64]. The phage cocktail comprised a five-phage mixture having each phage in the same ratio and was mixed with kasugamycin solutions diluted in nutrient broth at MIC, 1/2 MIC,1/4 MIC, and 0 MIC. The wild-type strain (10 5 CFU/mL) was co-cultured with a phage cocktail (10 8 PFU/mL) with or without a combination of antibiotics. The mixtures were cultured at 27 • C with shaking (150 rpm), and the cell counts were observed over time. Each experiment was performed in triplicates (n = 3).
Experiment on Apple Fruit under Controlled Conditions
Immature apples (cv. Fuji) were surface sterilized using ethanol, wounded, and infected with 2 × 10 5 CFU/mL of E. amylovora TS3128 according to a method described in a previous publication [66]. Wounded fruits were administered 2 × 10 8 PFU/mL of phages, antibiotics, or a phage-antibiotic combination and incubated in a humidified chamber at 27 • C. Symptoms were recorded at 2, 4, and 6 days after administration. The infected fruits were homogenized in order to enumerate the bacterial counts and the assay was repeated three times with three biological replicates (n = 3).
Stability Assay
The stability of the phages at different temperatures and pH conditions was examined. The phages (~1 × 10 8 PFU/mL) were incubated at 4 (control), 20, 30, 40, and 50 • C for thermal stability. The phages (~1 × 10 8 PFU/mL) were incubated in an SM buffer with a pH adjusted to 4, 5, 6, 7 (control), 8, and 9 using NaOH or HCl at 27 • C for the pH stability assay. After incubation for 60 min, the sample concentrations were evaluated in triplicates (n = 3). The stability value was standardized by using control as 100%.
DNA Isolation and Sequencing
The conventional phenol-chloroform method was used to isolate DNA from the phages [67]. RNase A (10 IU), DNase I (10 IU), and 10X DNase I buffer (Takara Bio, Kusatsu, Japan) were added to 1 mL of the phage solution of 10 10 PFU/mL, and then the solution was incubated at 37 • C for 1 h. Fifty microliters of 0.5 M ethylenediaminetetraacetic acid and proteinase K were added in the solution to inactivate the enzymes and hydrolyze the proteins, respectively. A mixture of isoamyl alcohol, chloroform, and phenol (1:24:25) was added, and the solution was centrifuged. Ethanol was added to the solution and the supernatant was removed. The precipitate was then resuspended in distilled water. The phage DNA was sequenced using an ABI 3730xl System (Thermo Fisher Scientific, Waltham, MA, USA) at Macrogen (Seoul, South Korea). FastQC (v0.11.6) was used to check the read quality. Trimmomatic (v0.36) was used to remove adapter sequences, and the assembly was performed using SPAdes (v3.12).
Statistical Analysis
Each experimental set of data of in vitro bacterial killing assay, phage-antibiotic synergy assay, and experiments on apple fruit under controlled conditions was statistically analyzed with one-way analysis of variance (ANOVA) and the Tukey post-hoc test using SigmaPlot software version 12.5 (Systat Software, San Jose, CA, USA).
Supplementary Materials:
The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/antibiotics11111566/s1, Figure S1: Host range of Erwinia phages ϕ27, ϕ31, ϕ32, ϕ47, and ϕ48, Figure S2: Stability of the phage virions, TableS1: Screening assay of the bacteriophages, Table S2: Functional classification of ORFs in Erwinia phage pEa_SNUABM_27, Table S3: Functional classification of ORFs in Erwinia phage pEa_SNUABM_31, Table S4: Functional classification of ORFs in Erwinia phage pEa_SNUABM_32, Table S5: Functional classification of ORFs in Erwinia phage pEa_SNUABM_47, Table S6: Functional classification of ORFs in Erwinia phage pEa_SNUABM_48. Data Availability Statement: All data generated or analyzed during this study are included in this published article.
Conflicts of Interest:
The authors declare no conflict of interest. | 8,632.4 | 2022-11-01T00:00:00.000 | [
"Biology"
] |
Identification of HPV16 in suspected cases of cervical lesions and docking Study of its L1 protein with some natural inhibitors.
Background: The Human papillomavirus (HPV) causes sexually transmitted diseases. Among several types of HPV variants, HPV 16 is listed as a high-risk group, the primary cervical cancer etiologic agent, which causes life-threatening disease among women worldwide. The presence of L1, E6 and E7 encoded oncoproteins are largely responsible for virulence and pathogenicity that leads to cervical lesions. This menace is required to be curbed by designing an anti-cancerous drugs. The protein receptor-inhibitor interaction adopted using in silico analysis is very important in drug designing. It was the objective of this study to identify HPV16 isolates from suspected cases of cervical cancer at SH Sokoto and SYMH Birnin Kebbi hospitals and also to identify potent HPV16’s L1 protein inhibitor using in silico analysis of Echinacoside, curcumin and Cichoric acid against the viral protein. Methods: A total of 140 cervical smear samples consisting of 21 low grade squamous intraepithelial lesion, 6 high grade lesion and 117 negative pap smears were collected. The samples were subjected for molecular detection using PCR targeting E6 and L1 genes of the virus. Positive samples were sequenced using Sanger sequencing platform. All the sequencing data were analysed using bioedit software while data generated for the molecular prevalence was statistically analyzed using Chi-square. A comprehensive HPV L1 protein homology model was designed to predict the L1 protein interaction mechanism with natural inhibitory molecules using a structural drug design approach. AutoDock Vina was used to carry out the molecular docking. Results: Out of the 140 samples, 24 samples were positive for the PCR representing 16.7% molecular prevalence rate. There is statistically significant association between cyto-diagnoses and presence of HPV16 ( P ˂0.05). The highest prevalence rate of 12(50% of positive sample) was recorded among women between 30-39 years old. Docking analysis showed that the Chicoric acid components of Echinacea purpurae have strong binding affinity to the L1 protein of the HPV. Conclusion: This study provides data on HPV 16 epidemiology in northern Nigeria, High-risk type 16 HPV variant was identified and also provides novel evidence for consideration on certain interacting residues, when synthesizing Anti-HPV compounds in the wet lab.
Background
Human papillomavirus is among the most prevalent infections transmitted during sexual activities which affects both male and female gender [1]. The size of the HPV genome is about 8kb, the genome is chromatinised, double stranded DNA (Deoxyribonucleic acid) which is enclosed in a 55 nm icosahedral capsid. HPV capsid is made up of two proteins, the L1 protein which is the major capsid protein and the L2 protein (the minor capsid protein). Each of the capsid consist of 360 monomers, which are arranged into 72 pentamers also known as capsomeres [2]. At early of HPV infection, the virion will first attach to heparin sulfate proteoglycans (HSPGs) situated in the cell surface or extracellular matrix [3]. L1 proteins will engage with HSPGs to cause conformational change in L1 and L2 proteins [4]. The host cell cyclophillin B facilitates the N-terminus to be expose, which will reveal a furin convertase cleavage site [5]. Following the conformational changes, the virion is associated with some non-HSPG receptors such as tetrasparins, annexin A2, growth factor receptor and integrins, to facilitate viral entry [6].
The virion enters the host through the endocytic route [6]. Human Papillomaviruses disease progression ranges from benign lesions to malignant [7]. The HPV types which are carcinogenic to the mucos membrane belongs to the genus alpha-papillomavirus, HPV types (HPV16 and 18), are the major cause of cancer of the cervix [8], they also cause vaginal, anal, oral, vulvar and penile cancers. The genus Alpha-papillomavirus also contains the benign mucosal HPV types that causes genital condylomas that are benign [9].
In developing countries HPV prevalence is higher, where asymptomatic infection is responsible for 44% of population [1]. A research was conducted in rural Nigeria, where 14.7 per cent of detectable high-risk HPV was registered, and representing two-thirds of the participants [10]. In northeastern Nigeria, 48.7% of the participant are positive for HPV type 18, while HPV type 16 was 13.2%, and HPV type 31, 33 and 35 accounted for 18.5% together [11]. Making HPV type 18 the predominant HPV in North-eastern Nigeria. Also, the high risk HPV is predominant in South-western part of Nigeria [12]. In kano North-west of Nigeria about 76% of the women positive for HPV are either HPV type 16, 18 or both while 60.5% are co-infected with HPV 16 and 18 [13].
The infections caused by HPV could lead to an end terminal stage disease, so the availability of some certain medicinal plants which possess pharmacological potentials are implored as chemotherapeutic agent for infections. In order to explore natural drugs with lesser side effect and cost, a natural compound Purple coneflower, is one among the plant reported to have active components such as chicoric acid, polysaccharides and echinacoside. This plant extract is known to stimulate immune response. The aqueous fractions of the stems, leaves, and flowers of Echinacea purpurea possess potent antiviral activity against HSV (Herpes Simplex Virus) type1 and 2, and hemagglutinin of influenza virus. This activity was attributed to the plant extract components, polysaccharide and cichoric acid. A potent antiviral photosensitiser was seen in the ethyl acetate and ethanol soluble fractions of the plants stem and leaves [14]. Another molecular docking research in which one of the plants components, L-chicoric acid was docked against the protein HIV-1 (Human immunodeficiency virus type 1) integrase [15], shows a very good binding modes between the ligand and the viral integrase, this explains its reported potency which is consistent with the experimental data available [15].
In most part of northern Nigeria, the prevalence of HPV is unknown, and the infection have been poorly managed in the region in question, therefore the main objective of this study to identify HPV16 isolates from suspected cases of cervical lesion at Specialist Hospital (SH) Sokoto and Sir Yahaya Memorial Hospital (SYMH) Birnin Kebbi and also to identify potential HPV16's L1 protein inhibitor using in silico analysis of Echinacoside, curcumin and Cichoric acid against the viral protein.
Study population
A total of 144 women with an average age of 33 years (range 20-60 years) participated in this study, population was recruited from patients attending Obstetrics and gynaecology units of SYMH Birnin Kebbi and SH Sokoto, Nigeria.
Sample collection
Cervical smear was collected from each of the participant by an application of standard procedure using sterilized speculum and swab (cyto-brush). The speculum was inserted through the vaginal orifice to allow visualization of the cervix. Cyto-brush was therefore placed into the endocervix and rotate in a circular fashion allowing collection of cellular smears from both ectocervix and endocervix (Squamocolumnar junction). The cyto-brush was removed, and its bristle was detached into the vial containing the fixative [16].
Nomenclature
This study adopted the following cytological classification: Negative; for normal cytology, low grade intraepithelial lesion, high grade intraepithelial lesion and carcinoma in situ; for malignancy.
DNA extraction
DNA from cervical samples were extracted using Viral Nucleic Acid Extraction kit II (Geneaid Biotech LTD, Taiwan). Extraction was carried out following the manufacturers' instruction.
Polymerase Chain Reaction
The PCR was conducted using KOD-FX Neo(Toyobo, Japan) following manufacturer instructions as follows, each PCR mix of 50µL contained, buffer 25µL, 1.3 µL of forward and reverse primer (HPV16 Pr1 and HPV16 Pr2) and each, Deoxynucleoside triphosphate (DNTP) of 10µL, molecular grade water of 10 µL, DNA template 1.4 µL, KOD of 1 µL. the following PCR conditions were used, 95°C for 5 minutes, followed by 35 cycles of 94°C for 30 seconds, 60°C for 30 seconds, 72°C for 30 seconds, final extension at 72°C for 5 minutes, then held for 16°C
Sequencing
The amplicons that showed desired band size at expected region for the target gene were sequence using sanger sequence method, sequences were aligned using BioEdit, and blasted on NCBI website, using the highly similar web tool.
Statistical analysis
Statistics analyses were performed using Chi-square with aid of an SPSS version 16.0, and data were presented in Tables and figures.
Hardware and software
All computational simulation studies were performed using (Intel Core i5-2430M) 6.00GB RAM with processor 2.40 GHz on windows7 operating system. Docking analysis was carried out using bioinformatics software such as PyRx virtual screening software (AutoDock Vina) and the visualization of structure was carried out using Pymol molecular graphic system with the procedure summarised in Figure 1. Online resources were also used in this study.
Ligand
For this present study, ligands were downloaded from zinc12 (zinc.org), and saved in mol2 format, and subsequently converted into pdbqt format by autodock vina.
Protein
Pentamer structure of L1 protein of human papillomavirus (PDB Code: 2r5h) was retrieved from the protein data bank and saved. The file was consequently opened by double clicking the folder containing the protein structure which comes out in Pymol. The receptor was prepared by deleting protein and water via edit then delete. The protein was converted from the downloaded file format to pdbqt format which is the Vina format.
Protein structure prediction
Phyre 2 server was used in the prediction of the L1 protein of the human papillomavirus.
Receptor-Ligand docking using PyRx virtual screening software
AutoDock Vina was used to carry out the protein-lıgand docking [17]. In order to perform protein-ligand docking, both the ligands and receptor (2r5h) were converted from pdb files to pdbqt (protein data bank, partial charge Q and atom type T) files (Vina input file format). AutoDock Tools (ADT) was used to prepare L1 protein of the human papillomavirus by the complete addition of hydrogen atoms to the receptor's carbon atoms. Non-polar hydrogen were added to the docked ligands. For Lamarckıan genetic algorithm, a maximum number of 15 × 10 5 energy evaluations, 27,000 maximum generations, 0.02 gene mutation rate and 0.8 crossover rate were used [18]. Each ligands has had two hundred independent docking runs.
Results
A self-administered questionnaire was given to participants, information on their sexual history was obtained. Table 1 shows that majority of the participants 64.1% have one sexual partner. Table 2 All the subjects that were positive for HPV type 16 complained of either genital itching, PV discharge or both symptoms, with genital itching (12) been most occurring symptoms among HPV positive subjects. The P value is statistically significant (˂0.05) as shown in table 2.
Discussions
HPV type 16 has been the major cause of cervical lesion in Africa, it causes 49% of cervical cancer in the continent, higher than any other HPV serotype.
HPV type 16 is responsible for 70 percent of all cervical cancer alongside with type 18 [19]. Of the 144 participants recruited in this study 16.9% were positive for HPV type 16, which is slightly higher than the study carried out in Kano, Nigeria where they reported prevalence of 15.8% for HPV type 16, though their study recruited lesser subjects (50) than this study [13]. A study carried out in Lagos, Nigeria, which they reported HPV type 16 to be 46.9%, this is not in accordance with the present study [20], another study reported 23.5% [21], while in a study carried out in the south western Nıgeria HPV 16 prevalence (3.5%) was very low than the one in this study [22]. study also showed HPV 16 is a strong risk factor for cervical lesion [21].
Symptoms such as genital itching, vaginal discharge (also known PV discharge), post-coital bleeding, are the most common symptoms seen in women with cervical abnormalities and vaginal infections. There was no complain of post-coital bleeding among the participants in this study, but all the subjects that were positive for HPV type 16 complained of either genital itching, PV discharge or both symptoms, with genital itching (12) been most occurring symptoms among HPV positive subjects. The P value is statistically significant (˂0.05). This is inconsistent with a study were they reported no association between infection with high risk HPV types and virginal problems (Itching, odor, and discharge) with P value of 0.14 [24]. though the symptoms can also be seen in other vaginal infections which are not tested for in this study, this might be the reason for the reported data not conforming with that of this study.
Conclusion
Numerous researches are currently ongoing in order to identify promising therapeutic agent for the management of HPV associated diseases, advancement in molecular modelling and bioinformatics are very important in validating those therapeutic agents using in silico analysis. This study provides data on some interacting residues for consideration, when synthesizing Anti-HPV compounds in the wet lab. Committee with reference number: SHS/SUB133/vol.1. The subjects whose signatures were obtained from the informed consent form participated in the | 2,952.6 | 2020-02-12T00:00:00.000 | [
"Medicine",
"Chemistry"
] |
Editorial for the Special Issue on Wide Bandgap Based Devices: Design, Fabrication and Applications, Volume II
Wide bandgap (WBG) semiconductors are becoming a key enabling technology for several strategic fields of human activities [...].
A power conditioning system is designed and built using SiC MOSFETs as switching devices by Ma et al. in [8], which, by leveraging the excellent thermal and voltage capability of SiC MOSFETs, is suitable for grid-level energy storage systems based on vanadium redox flow batteries. A digitally controlled photovoltaic emulator based on an advanced GaN power converter is developed by Ma et al. in [10], whereas in [9], the driving requirements of SiC MOSFETs and GaN HEMTs are illustrated, and the driving circuits designed for WBG switching devices are surveyed.
In [11], Kim et al. demonstrate that InGaN/GaN MQW LEDs on Si substrates with an AlN buffer layer grown with NH 3 interruption show improved crystal quality and enhanced optical output compared to LEDs with conventional AlN buffer. On the sensing application side, AlN is exploited by Chiu et al. [12] to fabricate piezoelectric micromachined ultrasonic transducers that are used to build a high-accuracy time-of-flight ranging system. Nguyen et al. [13] investigate the sensing characteristics of NO 2 gas sensors based on Pd-AlGaN/GaN HEMTs at high temperatures, while Thalhammer et al. [14] describe a novel class of X-ray sensors based on AlGaN/GaN HEMTs offering superior sensitivity and the opportunity for dose reduction in medical applications.
On the advanced processing technique side, laser micromachining on the frontside of SiC and sapphire wafers and the conditions by which the degradation of the performance of GaN HEMT electronics on the backside can be avoided are investigated by Indrišiūnas et al. in [15]. A novel dual laser beam asynchronous dicing method is proposed by Zhang et al. in [16] to improve the cutting quality of SiC wafers.
Regarding the properties and growth of emerging WBG materials, a methodology to synthesize gallium nitride nanoparticles by combining crystal growth with thermal vacuum evaporation is proposed by Fathy et al. in [22]. AlN is explored as an ultra WBG material in three papers: annealing Ni/AlN/SiC Schottky barrier diodes in an atmosphere of nitrogen and oxygen is shown to lead to a significant improvement in the electrical properties of the structures by Kim et al. in [19]; the effect of high-temperature nitridation and a buffer layer on semi-polar AlN films grown on sapphire by hydride vapor phase epitaxy is studied by Zhang et al. [21]; and the thermal annealing of AlN films with different polarities and its impact on crystal quality are studied by in Yue et al. in [23]. The effect of the annealing temperature on the microstructure and performance of sol-gel-prepared NiO films for electrochromic applications is analyzed by Shi et al. in [17]. Solution-processed In 2 O 3 thin films and TFTs are fabricated, and the factors affecting the stability of these devices are investigated by Yao et al. in [18]. The electronic structure and the optical properties of Srdoped β-Ga 2 O 3 are studied by Kean Ping et al. [20] using DFT first-principles calculations.
I would like to take this opportunity to thank all the authors for submitting their manuscripts to this Special Issue and all the reviewers for their time and their fundamental help in improving the quality of the accepted papers.
Conflicts of Interest:
The author declares no conflict of interest. | 789.6 | 2022-03-01T00:00:00.000 | [
"Engineering",
"Materials Science",
"Physics"
] |
The analysis of volatility of gold coin price fluctuations in Iran using ARCH & VAR models
Article history: Received June 28, 2013 Received in revised format 19 October 2013 Accepted 2 January 2014 Available online January 4 2014 The aim of this study is to investigate the changes in gold price and modeling of its return volatility and conditional variance model. The study gathers daily prices of gold coins as the dependent variable and the price of gold in world market, the price of oil in OPEC, exchange rate USD to IRR and index of Tehran Stock Exchange from March 2007 to July 2013 and using ARCH family models and VAR methods, the study analysis the data. The study first examines whether the data are stationary or not and then it reviews the household stability, Arch and Garch models. The proposed study investigates the causality among variables, selects different factors, which could be blamed of uncertainty in the coin return. The results indicate that the effect of sudden changes of standard deviation and after a 14-day period disappears and gold price goes back to its initial position. In addition, in this study we observe the so-called leverage effect in Iran’s Gold coin market, which means the good news leads to more volatility in futures market than bad news in an equal size. Finally, the result of analysis of variance implies that in the short-term, a large percentage change in uncertainty of the coin return is due to changes in the same factors and volatility of stock returns in the medium term, global gold output, oil price and exchange rate fluctuation to some extent will show the impact. In the long run, the effects of parameters are more evident. 2014 Growing Science Ltd. All rights reserved.
Introduction
The first decades of the new millennium has witnessed different world's challenges such as infamous September 11 incident, Iraq and Afghanistan war, etc.These incidents have created tremendous uncertainties and many investors moved to safe side in order to protect their investment, switching to gold from stock exchange.The world gold price went up from 600$ level to a historical record of 1900$ (Lawrence, 2003).Melvin and Sultan (1990) investigated South African political unrest, oil prices, and the time varying risk premium in the gold futures market and reported that while gold prices did not have any relationship with oil price but fluctuations on oil and stock exchange both influence on gold price, significantly.Bollerslev (1986) presented a natural generalization of the Autoregressive Conditional Heteroskedastic (ARCH) process to show for past conditional variances in the current conditional variance equation is proposed.Cai et al. (2001) provided a comprehensive characterization of the intraday return volatility in gold futures contracts traded on the COMEX division of the New York Mercantile Exchange.They detected employment reports, gross domestic product, consumer price index, and personal income as having the biggest impact.They also detected that the high-frequency returns disclosed long-memory volatility dependencies in the gold market, which had important implications on the pricing of long-term gold options and the determination of optimal hedge ratios.Tully and Lucey (2007) investigates macroeconomic influences on gold using the asymmetric power GARCH model (APGARCH) of Ding et al. (1993).They investigated both cash and futures prices of gold and substantial economic variables over the period 1983-2003, with special concentration on two periods, around the 1987 and 2001 equity market crashes.Their results indicated that APGARCH model could provide the most sufficient description for the data, with the inclusion of a GARCH term, free power term and unrestricted leverage impact term.Glosten et al. (1993) detected some support for a negative relationship between conditional expected monthly return and conditional variance of monthly return, using a GARCH-M model.Using the modified GARCH-M model, they also demonstrated that monthly conditional volatility could not be as persistent as was thought.Positive unanticipated returns seemed to result in a downward revision of the conditional volatility whereas negative unanticipated returns yield in an upward revision of conditional volatility.Ivanova and Ausloos (1999) presented a forecast of the low q-moment values of the assumed multifractal spectrum of Gold price, Dow Jones Industrial Average (DJIA) and Bulgarian Lev -USA Dollar (BGL-USD) exchange rate.The analysis demonstrated that these three financial data were not likely fractal but rather multifractal indeed.
The proposed model
The aim of this study is to investigate the changes in gold price and modeling of its return volatility and conditional variance model.There are two main hypotheses associated with the proposed study of this paper as follows, 1.The change on gold coin is a function of macro-economic factors and there are some meaningful relationships among them.2. The change on gold coin is a function of micro-economic factors and there are some meaningful relationships among them.
The study also considers whether there is some causality among various factors and whether the effects of positive and negative pulses are equal or not.The study gathers daily prices of gold coins as the dependent variable and the price of gold in world market, the price of oil in OPEC, exchange rate USD to IRR and index of Tehran Stock Exchange from March 2007 to July 2013 and using ARCH family models and VAR methods, the study analysis the data.The study first examines whether the data are stationary or not and then it reviews the household stability, ARCH and GARCH models (Engle et al., 1987;Engle & Kroner, 1995).
The effects of TARCH model on gold price
The proposed study investigates the causality among variables, selects different factors, which could be blamed of uncertainty in the coin return.Table 1 demonstrates the summary of some basic statistics.
Table 1
The summary of some basic statistics on gold price In this paper, we have performed Dickey Fuller test to see whether the data are stationary or not and Table 2 demonstrates the results of our investigation on price of gold (PCOIN).
The results of Table 2 clearly specify that Y is a stationary variable.The proposed study of this paper uses ARCH method with GARCH(1,1) as follows, Table 3 shows details of our results of the TARCH model for gold price.
Table 3
The results of TARCH model on gold price
The effect of TARCH model on oil price
Table 4 demonstrates the summary of some basic statistics.In addition, we have performed Dickey Fuller test to see whether the oil prices are stationary or not and Table 5 demonstrates the results of our investigation on price of oil (PROIL).
Table 6 shows details of our results of the TARCH model for oil price.
Table 6
The results of TARCH model on oil price
The effects of nonlinear EGARCH model on currency
The proposed study investigates the causality among variables, selects different factors, which could be blamed of uncertainty in the coin return.Table 7 demonstrates the summary of some basic statistics on currency data.Besides, we have performed Dickey Fuller test to see whether the oil prices are stationary or not and the proposed model find the following two equations as appropriate models, Table 8 shows details of our results of the EGARCH model for currency changes.
Table 8
The results of EGARCH model on currency changes Finally, we have performed augmented Dickey Fuller (ADF) to verify whether time series of gold price, oil price and currency are stationary or not and Table 9 shows details of our findings.The results of Table 9 specify that all data are stationary when the level of significance is one or five percent.
The VAR method
In this section, we present details of the implementation of VAR method.The proposed method uses the following time series equation, Table 10 demonstrates the results to find the optimum number of Inertia.
Table 10
The results of regression analysis According to Table 10, the best lag is determined as one based on Schwartz Baysian criterion.In addition, Fig. 1 demonstrates the stability of the VAR method.Res pons e of GARCHCOIN to PGOP Res pons e of GARCHCOIN to GARCHOIL Res pons e of GARCHCOIN to GARCHDR2 Response to Cholesky One S.D. Innovations ± 2 S.E.
As we can observe from the results of Fig. 2, the changes on local currency, world gold price, oil price have instance effects on gold coin.Next, we present details of analysis of variance for fluctuation of gold coin prices in different periods.Table 11 summarizes the results of our findings.As we can observe from the results of Table 11, during the first period, nearly all changes on gold coin fluctuations are associated with the gold coin price itself and world gold price as well as oil price did not influence on gold price, significantly.However, in other periods, other parameters such as world gold price, currency de-evaluation and stock exchange start influencing the gold price.
Conclusion
In this paper, we have presented an empirical investigation to study the effects of different factors such as world gold price, stock exchange, oil price and currency exchange on Iranian gold coin price.
The proposed study has gathered the historical information from March 2007 to July 2013 and using ARCH family models and VAR methods, the study analysis the data.The results have indicated that the effect of sudden changes of standard deviation and after a 14-day period disappears and gold price goes back to its initial position.In addition, in this study we have observed the so-called leverage effect in Iran's Gold coin market, which means the good news leads to more volatility in futures market than bad news in an equal size.Finally, the result of analysis of variance implied that in the short-term, a large percentage change in uncertainty of the coin return was due to changes in the same factors and volatility of stock returns in the medium term, global gold output, oil price and exchange rate fluctuation to some extent will show the impact.In the long run, the effects of parameters are more evident.
Fig. 1 .Fig. 2 .
Fig. 1.The results of inverse roots of AR characteristics polynomialThe results of Fig.1clearly show that the VAR model preserve sufficient stability.We now consider the effects of a shock on price of gold and these effects are shown in Fig.2as follows,
Table 2
The summary of Dickey Fuller test
Table 4
The summary of some basic statistics on oil price
Table 5
The summary of Dickey Fuller testThe results of Table5indicate that oil data become stationary after taking one difference.Here TARCH model for oil price is studied through the following relationship.
Table 7
The summary of some basic statistics on currency
Table 9
The results ADF test
Table 11
The summary of analysis of variance | 2,451.4 | 2014-01-01T00:00:00.000 | [
"Economics",
"Business"
] |
Improving Students' Understanding of Quantum Measurement Part 1: Investigation of Difficulties
We describe the difficulties that advanced undergraduate and graduate students have with quantum measurement within the standard interpretation of quantum mechanics. We explore the possible origins of these difficulties by analyzing student responses to questions from both surveys and interviews. Results from this research are applied to develop research-based learning tutorials to improve students' understanding of quantum measurement.
I. INTRODUCTION
Quantum mechanics is a particularly challenging subject, even for the advanced students [1][2][3][4].These difficulties have been described in a number of investigations [5][6][7][8][9][10][11][12].Based on these findings, we are developing a set of research-based learning tools to reduce students' difficulties and help them develop a solid grasp of quantum mechanics [13][14][15].This paper (Part 1) is the first of two in which we discuss the investigation of students' difficulties with quantum measurement.This investigation was conducted with the undergraduate and graduate students at the University of Pittsburgh (Pitt) and other universities by administering written tests and by conducting in-depth individual interviews with a subset of students.The development of the research-based learning tools and the preliminary evaluation of students' performance after using the learning tools is described in the second of the two papers (Part 2) [16].
The standard formalism of quantum measurement (which is taught to the undergraduate and graduate students universally) is quite different from classical mechanics, where position and momentum of a particle evolve in a deterministic manner based upon the interactions [1].In quantum mechanics, position, momentum and other observables are in general not well-defined for a given state of a quantum system.The time-dependent Schroeinger equation (TDSE) governs the time evolution of the state which can be written as a linear superposition of a complete set of eigenstates of any hermitian operator corresponding to a physical observable.The state of the system evolves in a deterministic manner depending on the Hamiltonian of the system.According to the Copen-hagen interpretation, quantum measurement would instantaneously collapse the wavefunction (or the state of the system) to an eigenstate of the operator corresponding to the physical observable measured and the measured value is the corresponding eigenvalue.For example, in an ideal measurement, if we measure the position of a quantum particle in a one-dimensional (1D) infinite square well, its wavefunction will collapse to a position eigenfunction which is a delta function in the position representation.If we measure its energy instead, the wavefunction of the system will collapse into an energy eigenfunction, which is a sinusoidal function inside the 1D well and goes to zero at the two boundaries (and is zero everywhere outside the well).
The eigenvalue spectrum of an operator can either be discrete or continuous or a combination of the two.In an N dimensional Hilbert space, an operator Q corresponding to a physical observable Q with discrete eigenvalue spectrum has N eigenvalues q n and corresponding eigenstates |q n .
The state of the system at a given time, |Ψ(t) , can be written as a linear superposition of a complete set of eigenstates of |q n .By projecting the wavefunction of the system |Ψ(t) at time t onto an eigenstate |q n of the operator Q , we can find the probability | q n | Ψ(t) | 2 of obtaining q n when the observable Q is measured at time t.
After the measurement of the observable Q, the time-evolution of the state of the system, which is an eigenstate of Q right after the measurement, is again governed by the TDSE.Right after the measurement of energy, the state of the system collapses into the same energy eigenstate, and the probability density does not change with time (we only focus on time-independent Hamiltonians) since the only change in the wavefunction with time is an overall time-dependent phase factor.If the system is initially in an energy eigenstate at time t = 0 and we measure an arbitrary physical observable Q after a time t, the probability of obtaining an eigenvalue will be time-independent since the system was still in an energy eigenstate at time t at the instant the measurement of Q was performed.Therefore, the energy eigenstates are called the stationary states.On the other hand, a measurement of position would collapse the system into a position eigenstate at the instant the measurement is made.However, a position eigenstate is a linear superposition of the energy eigenstates and the different energy eigenstates in the linear superposition will evolve with different time-dependent phase factors.Therefore, the probability density after position measurement will change with time.In this case, the probability of measuring a particular value of energy will be time-independent but the probability of measuring another physical observable whose operator does not commute with the Hamiltonian will depend on time.
The goal of the investigation was to examine students' difficulties with quantum measurement after traditional instruction.The topics included in the investigation, such as the measurement outcomes, probability of obtaining an eigenvalue, stationary states and eigenstates, etc., were all covered in the traditional instruction of quantum mechanics.The investigation was carried out over several years.For example, students were given the questions as part of the concept tests, quizzes or tests depending upon the instructor's preference.Therefore, the number of students who answered a particular question varies.To simplify the mathematics and focus on the concepts related to quantum measurement, we often used the model of a 1D infinite square well during the investigation.Both open-ended questions and multiple-choice questions were administered to probe students' difficulties.We also had informal discussions with a subset of students who took the written test and formally interviewed some students to get a better understanding of students' reasoning process.
Different Observables
The measurement of a physical observable collapses the wavefunction of the quantum system into an eigenstate of the corresponding operator.Many students have difficulties distinguishing between energy eigenstates and the eigenstates of other physical observables.To investigate the pervasiveness of this difficulty in distinguishing between the eigenstates of different physical observables, one of the multiple choice questions administered to the students was the following: • Choose all of the following statements that are correct: (1) The stationary states refer to the eigenstates of any operator corresponding to a physical observable.
(2) If a system is in an eigenstate of any operator that corresponds to a physical observable, it stays in that state unless an external perturbation is applied.
(3) If a system is in an energy eigenstate at time t = 0, it stays in the energy eigenstate unless an external perturbation is applied.1.The most common incorrect choice was E (all of the above).Nearly half of the students thought that all three statements were correct because they had difficulty in differentiating between the related concepts of stationary states and eigenstates of other observables.Some students selected choice A (1 only) which is interesting because one may expect that students who claimed statement (1) was correct and understood why a stationary state is called so may think that statement (2) is correct as well.In particular, for students who claimed statement (1) is correct, statement (2) may be considered "a system in a stationary state stays in that state unless an external perturbation is applied", which described the property of stationary state.However, students who selected choice A did not relate the stationary state with the special nature of the time evolution in that state.
B. Difficulty with possible outcomes of a measurement and the expectation value of the measurement result
The following is an example of a multiple choice question which was administered to investigate students' understanding of the possible outcomes of a measurement for a given state of a particle in a 1D infinite square well when the measurement is performed.ψ 1 (x) and ψ 2 (x) are the ground state and first excited state wavefunctions.
• An electron is in the state given by ψ 1 (x)+ψ 2 (x) √ 2 . Which one of the following outcomes could you obtain if you measure the energy of the electron?
A. Informal discussions with individual students and individual think-aloud interviews [17] indicated that many students were not only confused about the distinction between individual measurements and expectation values, they also had difficulty distinguishing between the probability of measuring a particular value of an observable in a given state and the measured value or the expectation value.
For example, during individual interviews, students often wrote as the probability of measuring E n in the state |Ψ .When these students were explicitly asked to compare their expressions for the probability of measuring a particular value of energy and the expectation value of energy, some students appeared concerned.They recognized that these two concepts were different but they still struggled to distinguish these concepts.They could not write an expression for the probability of measuring E n either using the Dirac notation or in the position space representation using the integral form.
Discussions with students and individual interviews suggest that some of them had difficulty in differentiating between the probability of measuring each possible value of an observable and the expectation value of that observable in a given state.Since the expectation value in a given state equals the average of a large number of measurements of that observable on identically prepared systems, it is equal to the sum of the eigenvalues of the corresponding operator times their probabilities in the given state.Many students had difficulty with the statistical interpretation of the expectation value of Q as the average of a large number of measurements on identically prepared systems in state |Ψ .For example, a survey question which was administered to 202 graduate students from seven universities illustrates it as shown below [10]: • The wavefunction of an electron in a 1D infinite square well of width a at time t=0 is given by Ψ(x, t = 0) = 2/7ψ 1 (x) + 5/7ψ 2 (x) .Answer the following questions.(a) You measure the energy of an electron at time t=0.Write down the possible values of the energy and the probability of measuring each.
(b) Calculate the expectation value of the energy in the state Ψ(x, t) .
67% of the graduate students answered question (a) correctly and 7% of them were confused about the distinction between the energy eigenvalues and the expectation value of energy.
However, only 39% of the students provided the correct response for question (b) above.Many students who could calculate the probability for measuring each energy in question (b) did not use the probabilities to find the expectation value.Some of them tried to find the expectation value by sandwiching the Hamiltonian with the state of the system (i.e., Ψ| Ĥ |Ψ ) which is correct and some even wrote down correct corresponding integrals but then struggled with the calculation.
C. Difficulty with the probability of measuring energy
When we explicitly asked students to find the probability of obtaining energy E 2 for the in a 1D infinite square well, many of them could provide the correct answer 1/2 by observing the coefficients.To evaluate whether students could calculate the probability of measuring a particular value of energy by projecting the state vector along the corresponding energy eigenstate for the case where the wave function is not written explicitly in terms of a linear superposition of energy eigenstates, the following question about a triangle shaped wavefunction in a 1D infinite square well was administered: • The state of an electron at t=0 is given by x < a and Ψ(x) = 0 elsewhere.Here A is the normalization constant.What is the probability that an energy measurement at time t=0 yields energy E 2 ?(If there is an integral in your expression for the probability, you need not evaluate the integral but set it up properly with appropriate limits.Ignore the fact that the first derivative of the wavefunction is not continuous.)A similar multiple-choice question about a parabola shaped wavefunction was administered to 76 students in six universities as shown below: • Consider the following wavefunction for a 1D infinite square well: Ψ(x) = Ax(a − x) for 0 ≤ x ≤ a and Ψ(x) = 0 otherwise.A is a normalization constant.Which one of the following expressions correctly represents the probability of measuring the energy E n for the state Ψ(x) ?
Among the 76 students, 61 were junior/senior undergraduate students and the others were first year graduate students in physics department.The distributions of the undergraduate and graduate students' answers are listed in Table 3.About one third of both the graduate and undergraduate students chose the correct answer B. The undergraduate students tended to include the Hamiltonian operator in calculating the probability of measuring a particular energy eigenvalue.For example, 49% of the undergraduate students incorrectly selected the distractor option A which is and E 2 could also be obtained but the probability of measuring E 1 would be largest.Another 12% of the students thought that all the possible energies E n can be measured with the same probability.
D. Difficulties with the time development of the wavefunction after the measurement of an observable
Within the Copenhagen interpretation of quantum mechanics, the measurement of an observable is treated separately from the "normal" time-evolution of the system according to TDSE.When a measurement is performed, the state of the system instantaneously collapses to an eigenstate of the operator corresponding to the observable measured after which the system will evolve normally according to the TDSE.We investigated students' understanding of the time-development of the wavefunction according to the TDSE after the measurement of an observable by asking 15 students the following question about consecutive position measurement for a 1D infinite square well: • If you make a measurement of position on an electron in the ground state of a 1D infinite square well and wait for a long time before making a second measurement of position, do you expect the outcome to be the same in the two measurements?Explain.
To correctly answer this question, students must know the following: (1) the ground state wavefunction will collapse into a position eigenfunction (a delta function in position) after the first position measurement; (2) the position eigenfunction is not a stationary state wavefunction so the wavefunction will evolve in time in a non-trivial manner and it will not in general be found in a position eigenstate after a time t.Therefore, after a long time, the second measurement of position in general will yield a different value from the first measurement.We note however that in an infinite square well, the time evolution of the system is such that the wave function repeats itself with a certain periodicity.
Difficulity D.1: System remains in the energy eigenstate after a position measurement In response to this question, some students thought that the system will be in the ground state after both the first and the second position measurements.Informal discussions with some students and formal interviews with a handful of students suggest that those with these types of responses often did not realize the difference between an energy eigenstate and a position eigenstate.They claimed that if the system is in the ground state, it will remain in that state.
Students who were explicitly asked what would happen if the initial state before the measurement was the first excited state (which is also an energy eigenstate) typically responded that it will remain in that state since even that state is an "eigenstate".In the written survey, only one out of fifteen students explicitly mentioned the wavefunction collapse after the first position measurement.However, his response was "the wavefunction collapses into the measured state" and he did not elaborate that the "measured state" is actually a position eigenstate.
Difficulty D.2: System stays in the position eigenstate at any time after a position measurement Some students claimed that after the first position measurement the system gets "stuck" in a position eigenstate and did not know that the position eigenfunction (unlike the energy eigenfunctions for a time-independent Hamiltonian) evolves in time in a non-trivial manner and the system does not remain a position eigenfunction for all future time t.These students claimed that the second position measurement will yield the same value as the first one unless there was an "outside disturbance".Only two out of fifteen students mentioned the correct time evolution of the quantum mechanical system after the position measurement.
Difficulty D.3: System finally goes back to the initial state
Students were also asked another series of questions about measurement when the initial state of the system at time t = 0 is Ψ(x, 0) = 2/7ψ 1 (x) + 5/7ψ 2 (x) for an electron confined in a 1D infinite square well as follows: • Q1.If the energy measurement yields 4π 2 h2 /(2ma 2 ), what is the wavefunction right after the measurement?
• Q2.Immediately after the energy measurement in Q1, you measure the position of the electron.What possible values could you obtain and what is the probability of each?
• Q3.After the position measurement in Q2, you wait for time t > 0 and measure the position again.Would the probability of measuring each possible value different from Q2?
Q1 which asks about the state of the system long after the energy measurement (instead of immediately after the measurement as in the open-end question) has been given as a multiple-choice question to 76 students from 6 universities.An analysis of the student responses suggests that 20% of the students did not know that the wavefunction would collapse at the instant the energy was measured.Also, 36% of the students thought the wavefunction will collapse upon energy measurement but finally evolved back to the initial state 2/7ψ 1 (x) + 5/7ψ 2 (x) long time after the measurement.During the individual interview, a student said, "it's like tossing a coin.
You can get either head or tail after the measurement.But when you make another measurement, it goes back to a coin (with two sides)."Such a statement also indicates that the student made an inappropriate transfer of a classical probability concept to quantum probability.
Difficulty D.4: Probability density for position measurement
Born's probabilistic interpretation of the wavefunction can also be confusing for students.In Q2 above, the wavefunction of the system before the position measurement is the energy eigenstate ψ 2 (x) = 2/a sin(2πx/a).We expected students to note that one can measure position values between x = 0 and x = a (except x = 0, a/2, and a where the wavefunction is zero), and according to Born's interpretation, |ψ 2 | 2 dx gives the probability of finding the particle in a narrow range between x and x + dx.However, only 38% of students provided the correct response.Partial responses were considered correct for tallying purposes if students wrote anything that was correct related to the above wavefunction, e.g., "The probability of finding the electron is highest at a/4 and 3a/4", "The probability of finding the electron is non-zero only in the well", etc. [18] Eleven percent of the students tried to find the expectation value of position instead of the probability of finding the electron at a given position.They wrote the expectation value of position in terms of an integral involving the wavefunction.Many of them explicitly wrote that probability = (2/a) a 0 x sin 2 (2πx/a)dx and claimed that instead of the expectation value they were calculating the probability of measuring the position of the electron.
During the interview [18], one student said (and wrote on paper) that the probability of position measurement is x |ψ| 2 dx ( ψ = ψ 2 in Q2 above).When the interviewer asked why |ψ| 2 should be multiplied with x and if there is any significance of |ψ| 2 dx alone without multiplying it by x, the student said, "|ψ| 2 gives the probability of the wavefunction being at a given position and if you multiply it by x you get the probability of measuring (student's emphasis) the position x".
When the student was asked questions about the meaning of the "wavefunction being at a given position", and the purpose of the integral and its limits, the student was unsure.He said that the reason he wrote the integral is because x |ψ| 2 dx without an integral looked strange to him.
Similar confusion about probability in classical physics situations have been found [19].
Difficulty D.5: Use of classical language to describe time evolution of quantum systems Out of the ten students who were given Q3 above, none of them could answer it correctly though it assesses the same concepts as in the consecutive position measurement question discussed earlier.
In the consecutive position measurement question, some students used a classical description to answer the question about the time-evolution after the measurement such as "the electron moves around".Discussions with individual students and interviews suggest that such classical responses reflect students' discomfort describing the time evolution of a quantum system in terms of the time-development of wavefunction.
E. Incorrectly believing that an operator acting on a state corresponds to a measurement of the corresponding observable One of the questions on a survey given to more than 200 graduate students asked them to consider the following statement [10]: "By definition, the Hamiltonian acting on any allowed (possible) state of the system |ψ will give the same state back, i.e.Ĥ |ψ = E |ψ ,where E is the energy of the system."Students were asked to explain why they agree or disagree with this statement.We expected students to disagree with the statement and note that it is only true if |ψ is a stationary state.In general, |ψ = Eleven percent of the students answering this question incorrectly claimed that any statement involving a Hamiltonian operator acting on a state is a statement about the measurement of energy.Some of these students who incorrectly claimed that Ĥ |ψ = E |ψ is a statement about energy measurement agreed with the statement while others disagreed.Those who disagreed often claimed that Ĥ |ψ = E n |ψ n because as soon as Ĥ acts on |ψ , the wavefunction will collapse into one of the stationary states |ψ n and the corresponding energy E n will be measured.The following are two typical responses in this category: • Disagree.Hamiltonian acting on a state (measurement of energy) will return an energy eigenstate.
• When |ψ is a superposition state and Ĥ acts on |ψ , |ψ evolutes to one of the |ψ n so we have Ĥ |ψ = E n |ψ n .
Formal interviews, informal discussions and written reasonings suggest that these students often believed that the measurement of any physical observable in a particular state is achieved by acting with the corresponding operator on the state.The incorrect notions expressed above are often overgeneralizations of the fact that after the measurement of energy, the system is in a stationary state so Ĥ |ψ n = E n |ψ n and students felt that there should be an equation describing the collapse of the wave function.
Individual interviews related to this question suggest that some students believed that whenever an operator Q corresponding to a physical observable Q acts on any state |ψ , it will either yield a corresponding eigenvalue λ and the same state back, i.e., Q |ψ = λ |ψ or yield Q |ψ = λ n |φ n where |φ n is the n th eigenstate of Q in which the system collapses and λ n is the corresponding eigenvalue (but actually, Q |φ n = λ n |φ n ).
We further explored this issue by asking 17 and 15 graduate students at the end of their first semester and second semester graduate level quantum mechanics course the following question.15 graduate students were the same in both semesters.
• Consider the following conversation between Andy and Caroline about the measurement of an observable Q for a system in a state |ψ which is not an eigenstate of Q : Andy: When an operator Q corresponding to a physical observable Q acts on the state |ψ it corresponds to a measurement of that observable.Therefore, Q |ψ = q |ψ where q is the observed value.
Caroline: No.The measurement collapses the state so Q |ψ = q |ψ q where |ψ q on the right hand side of the equation is an eigenstate of Q with eigenvalue q.
With whom do you agree?E. The answer depends on the observable Q.
We note that the question was not posed as a multiple-choice question at the end of the first semester course but students were asked to explain whom if any they agreed with and why.There was a brief discussion of the correct response to the question after administering the survey in which this question was asked.At the end of the first semester course, 12% of the students agreed with Andy, 47% with Caroline, 29% with neither (correct response) and 12% provided no response.In the second semester, the concepts about measurement were not explicitly emphasized in the course of Quantum Mechanics II.At the end of the second semester course, the same question in the multiple-choice form was administered.This time, 20% of the students chose the distractor A, which is lower than the 47% in the first semester.However, about the same percentage of students in both semesters thought that the operator corresponding to an observable acting on any quantum state gives the eigenvalue, i.e., Q |ψ = q |ψ .The comparison of students' answer distribution is listed in Table 4. 13% of the students agreed with Andy, 20% with Caroline, 7% with both and 53% with neither (correct response).While the percentage of correct response increased significantly from the first to the second administration, many students still had difficulty with this concept.
Earlier, the version of this question not in the multiple-choice format was posed to 37 graduate students at the beginning of their graduate level quantum mechanics course (not the same students as those who answered it at the end of the first and second semester of their graduate level quantum mechanics course).In that group, 24% of the students agreed with Andy, 54% with Caroline and 22% with neither (correct response).Indeed this difficulty is quite common even amongst graduate students and graduate level instruction does not help students develop a better understanding of these concepts.
III . SUMMARY AND CONCLUSION
We find that students share common difficulties with concepts related to quantum measurement.
In particular, many students were unclear about the difference between energy eigenstates and eigenstates of other physical observables and what happens to the state of the system after the measurement of an observable.Students also had difficulty in distinguishing between the measured value, the probability of measuring it and the expectation value.They often did not think of the expectation value of an observable as an ensemble average of a large number of measurements on identically prepared systems but rather thought of it as a mathematical procedure where an operator is sandwiched between the same bra and ket states (the state of the system) or the integral formulation for calculating the expectation value in the position representation.Students were also confused about whether the system is stuck in the state in which it collapsed right after the measurement or whether it goes back to the state before the measurement was performed.
In general, students struggled with issues related to the time evolution of wave function after the measurement.Based on the investigation of students' difficulties, we developed the Quantum Interactive Learning Tutorial (QuILT) and concept tests to improve students' understanding of quantum measurement.These research-based learning tools will be discussed in the second of the two papers (Part 2).
A. 1
only B. 3 only C. 1 and 3 only D. 2 and 3 only E. all of the above The correct answer is B (3 only).In statement (1), the stationary states should refer to which is composed of only two energy eigenstates, the triangle function state |Ψ (or Ψ(x) in the position space) is a superposition of infinitely many energy eigenstates, i.e., |Ψ = ∞ n=1 c n |ψ n .The expansion coefficient equals ψ n |Ψ = +∞ −∞ ψ * n (x)Ψ(x)dx and |c n | 2 is the probability of obtaining energy E n when energy is measured for the state |Ψ .Thus, to answer this question correctly, students need to write |Ψ as a linear superposition of |ψ n and find the component of |Ψ along |ψ n .Only one student out of fifteen provided the correct answer and some students left this question blank.Other students had two common mistakes.Twenty percent of the students wrote down the energy expectation value Ψ| Ĥ |Ψ to represent the energy measurement probability.In further informal discussions and formal interviews with some students, we asked how the expression Ψ| Ĥ |Ψ which only involved state |Ψ would favor energy E 2 over any other energy.Some of the students then changed their answers to ψ 2 | Ĥ |Ψ which was still incorrect.Another 27% of the students claimed that the "probability" of measuring any physical observable was represented by |Ψ(x)| 2 according to the interpretation of wavefunction.These students were confusing the probability density for measuring position with the probability of measuring other physical observables such as energy.
TABLE I :
The choice distribution of 10 students answering the question about stationary state and eigenstate after traditional instruction.
Because the energy eigenstates |ψ n are orthogonal to each other, | ψ n | Ψ | 2 = 1/2 for both n = 1 and n = 2 and | ψ n | Ψ | 2 = 0 for all the other energy eigenstates E n (n > 2).Therefore, we can only obtain E 1 or E 2 with equal probability but no other energy.The distribution of students' answers is shown in Table 2. Six out of fifteen students in a junior-senior quantum mechanics class chose the correct answer C (either E 1 or E 2 ).The most common incorrect choice selected by 27% of the 15 students was B ( (E 1 + E 2 )/2 ) which actually represents the expectation value of energy.Students mistakenly claimed that the expectation value is the measured value of energy.
Either E 1 or E 2 D. Any of E n (n=1,2,3,) E. Any value between E 1 and E 2
TABLE II :
The choice distribution of 15 students answering the question about energy measurement outcome after traditional instruction.
TABLE III :
The choice distributions of 61 undergraduate students and 15 graduate students answering the question about energy measurement probability.Another multiple choice question given to the same 76 students asked about the energy measurement outcomes for the state 4/7 |ψ 1 + 3/7 |ψ 2 .55% of all the 76 students provided the correct answer.21% of the students incorrectly claimed that other energies E n besides E 1
TABLE IV :
The answer distributions of 17 graduate students in the first semester and 15 graduate students in the second semester answering the same question about energy measurement. | 7,176.4 | 2012-04-19T00:00:00.000 | [
"Physics"
] |
Tailoring oxide properties : an impact on adsorption characteristics of molecules and metals
Both density functional theory calculations and numerous experimental studies demonstrate a variety of unique features in metal supported oxide films and transition metal doped simple oxides, which are markedly different from their unmodified counterparts. This review highlights, from the computational perspective, recent literature on the properties of the above mentioned surfaces and how they adsorb and activate different species, support metal aggregates, and even catalyse reactions. The adsorption of Au atoms and clusters on metal-supported MgO films are reviewed together with the cluster’s theoretically predicted ability to activate and dissociate O2 at the AuMgO(100)/Ag(100) interface, as well as the impact of an interface vacancy to the binding of an Au atom. In contrast to a bulk MgO surface, an Au atom binds strongly on a metal-supported ultra-thin MgO film and becomes negatively charged. Similarly, Au clusters bind strongly on a supported MgO(100) film and are negatively charged favouring 2D planar structures. The adsorption of other metal atoms is briefly considered and compared to that of Au. Existing computational literature of adsorption and reactivity of simple molecules including O2, CO, NO2, and H2O on mainly metal-supported MgO(100) films is discussed. Chemical reactions such as CO oxidation and O2 dissociation are discussed on the bare thin MgO film and on selected Au clusters supported on MgO(100)/metal surfaces. The Au atoms at the perimeter of the cluster are responsible for catalytic activity and calculations predict that they facilitate dissociative adsorption of oxygen even at ambient conditions. The interaction of H2O with a flat and stepped Ag-supported MgO film is summarized and compared to bulk MgO. The computational results highlight spontaneous dissociation on MgO steps. Furthermore, the impact of water coverage on adsorption and dissociation is addressed. The modifications, such as oxygen vacancies and dopants, at the oxide-metal interface and their effect on the adsorption characteristics of water and Au are summarized. Finally, more limited computational literature on transition metal (TM) doped CaO(100) and MgO(100) surfaces is presented. Again, Au is used as a probe species. Similar to metal-supported MgO films, Au binds more strongly than on undoped CaO(100) and becomes negatively charged. The discussion focuses on rationalization of Au adsorption with the help of Born-Haber cycle, which reveals that the so called redox energy including the electron transfer from the dopant to the Au atom together with the simultaneous structural relaxation of lattice atoms is responsible for enhanced binding. In addition, adsorption energy dependence on the position and type of the dopant is summarized.
I. INTRODUCTION
Metal oxides have long been considered as potential materials for large variety of applications ranging from gas sensing and protective coatings to electrodes in fuel cells, heterogeneous photo(electro) catalyst, and bio-compatible materials 1 . Compared to any other material class both crystallographic and electronic properties of oxides display diverse behaviour, e.g., electronic conductivity ranges from wide-gap insulators to materials with conductivity comparable to metals 2 . The characteristics of oxides can be tailored to improve desired properties in various ways by introducing structural modifications like steps and grain boundaries, adding impurity atoms as dopants, or removing atoms from the structure 1,2 . In particular, point defects such as oxygen vacancies determine optical, electronic and transport properties of insulating oxides, and they usually dominate the chemistry of its surface 3 . While transition metal oxides are utilized for their catalytic properties, simple oxides such as MgO or CaO are intrinsically inert owing to their very deep valence band and very high conduction band; thus, they are less exploited in applications. However, simple oxides are interesting model systems whose properties have been thoroughly investigated 1,4 and therefore they form an ideal platform to explore the impact of different tailoring strategies to improve their reactivity. One way to achieve this is to prepare oxides as metal grown thin films, which provides an unique approach to modify structural, electronic and chemical properties as a function of film thickness extensively, which is discussed in the reviews of Prof. H.J. Freund, Prof. G. Pacchioni, and Prof. N. Nilius [5][6][7][8] . From the experimental point of view metal supported thin-film systems create specific technical challenges to be tackled. The significant benefit of ultra-thin arrangement of insulating oxides is that they can be studied with scanning tunneling microscopy (STM), which is not possible for their bulk counterparts.
Ultra-thin oxide films such as MgO 6,9 , NiO 10 , CaO 11 , Al 2 O 3 12,13 , FeO 14-16 , and SiO 2 17 have been extensively studied. A comprehensive overview of metal-supported transition metal oxide films can be found in reference 18 . Possible applications of metal supported ultra-thin oxide films can be divided into two groups: support materials and active players in chemical conversions. Gold clusters on metal supported ultra-thin films show distinct features compared to clusters on bulk films 5 and calculations predict that the perimeter of these particles is highly reactive e.g., activating oxygen readily 19 . Supported films can also directly act as a catalyst for CO oxidation 20,21 and dissociate H 2 22 . Furthermore, metal-supported ultra-thin 4 films are reactive towards water dissociation 23 .
Other ultra-thin insulating materials grown on metal surfaces include e.g., NaCl on Cu (111), which is used as a substrate to explore charge transport to nanostructures including Au adatoms [24][25][26] . While the stoichiometry and atomic structure of ultra-thin MgO corresponds that of the bulk MgO, this is not always the case. The most prominent example of this is an ultra-thin alumina film over a NiAl support for which determination of atomic structure turned out to be particularly challenging owing to a complex atomic structure.
From the interplay between STM experiments and density functional theory (DFT) calculations, the peculiar alumina structure was revealed and it corresponds to Al 10 However, dopant atoms can have either lower or higher valence compared to a cation they substitute and this dictates their impact on host oxide characteristics. In the latter case, the 5 influence of a dopant resembles that of a metal support in the ultra-thin oxide film. However, the number of spare electrons depends on the concentration of dopants and the fact that the dopants can be mobile, unlike the metal support. There is a large body of studies, where the impact of substutional doping on catalytic properties of oxides has been examined for example see references in 32,33 . In many cases it has been shown that the reactivity has improved upon doping. Among the studied systems Li-doped MgO has received the most attention owing to its use as a catalyst for oxidative methane coupling to ethane that is often labeled as Holy Grail in catalysis [34][35][36] . Despite extensive research on this topic, many questions have remained open, and there is not a conclusive understanding of the role of the catalyst, let alone its structure. In the case of CaO, STM studies demonstrate that tiny amounts of Mo embedded into the oxide introduce similar features as metal-supported ultra-thin MgO films namely, Au atoms become negatively charged 37,38 and Au aggregates favour the 2D growth mode 39 . Calculations predict enhanced binding of Au adatoms, which is due to electron transfer from Mo to the adsorbate and substantial lattice relaxations near the dopant originating from increased attraction between more positively charged Mo and surrounding anions 40 . Notwithstanding the similarities, TM-doped simple oxides and metal-supported ultra-thin oxide systems show fundamental differences as impact of dopant is more localised compared to metal support, and a number of electrons that the dopant can provide is very limited, which means that the dopant concentration is a new adjustable parameter.
II. PREPARATION AND PROPERTIES OF MODIFIED SIMPLE OXIDES
Here the preparation of oxide systems is briefly discussed and for a more detailed discussion on the preparation and experimental studies of ultra-thin films and simple doped oxides refer to previous reviews 5, 6,8 . Ultra-thin-oxide layers are synthesized by the vaporization of metals such as Mg in a background of molecular oxygen on a metal support and they typically have smaller optimal lattice parameters than bulk oxides 41 . By far the most extensively studied ultra-thin oxide film is MgO for which the best support is either a Ag (100) or Mo (100) surface owing to small lattice mismatch, which amounts to 3% and 5.4 %, for Ag and Mo respectively. In the thin film structure O anions are preferably located above Ag/Mo atoms 42 as displayed in Figure 1. Already a 3ML-thick MgO film on a Ag-support is sufficient to reproduce the band gap of bulk MgO 9 .
Crucial to the characteristics of metal-supported ultra-thin oxide materials is the interaction between a metal substrate and an oxide film, which affects both the geometric and electronic properties of the system. Furthermore, even a small lattice mismatch, like the In the pioneering work of Haruta and Hutchings, it was found that Au nanoparticles can catalyze CO oxidation to CO 2 at low temperature 75,77 and acetylene hydrochlorination 75,78,79 .
Later, Au nanoparticles were found to be active in reactions such as propylene epoxidation 80 , Calculations from Honkala and Häkkinen show that Au adsorption energy depends sensitively on film thickness being strongest on the thinnest film, Figure 2 with the exception that a 2ML-thick MgO film binds Au more strongly than a 1ML-thick film 48 image appears because tunneling is not in resonance with any cluster states. Adjusting the bias to appropriate values, flower-like shapes appear and the increase in cluster height is seen. Both effects suggest that the STM contrast is dictated by electronic structure of a cluster whose exact atomic structure remains unknown. Theoretically, the electronic structure of a planar cluster can be predicted employing the 2D harmonic oscillator potential model, which confines the atomic 6s valence electrons 118 . The eigenstates of the 2D harmonic oscillator (HO) can be characterised by their principal quantum number together with the projection of angular momentum on the normal vector of the plane, m. Thus, the states have n-1 radial nodes and |m| angular nodes. The atomic nomenclature of the states is adopted, which means that e.g., the 2P state has one radial node and two angular nodes.
Moreover, states such as 1F exhibit a very flower-like appearance. Again, each Au atom brings in one 6s electron, therefore the number of atoms in a cluster can be obtained by analyzing the nodal structure of frontier orbitals in a cluster within the 2D HO model. To resolve the atomic structure of these clusters, an extensive DFT search for possible structures was carried out at size regime of 6-20 Au atoms per a cluster. The aim was to find clusters having similar HOMO and LUMO states as seen in STM images of unknown species. The candidate cluster sizes were selected so that the 2D HO model predicts that their HOMO and LUMO states resemble those seen in experiments. Furthermore, since flower-like shapes are typical for symmetric planar clusters the computational search focused on these structures. Similar to 1D Au chains Au clusters might have additional electrons not originating from 6s states. Figure 8 collects some of the selected cluster sizes and shapes calculated. All of them experience thin-film effects that is having a planar geometry with formation energies ranging from -2.1 --2.5 eV/atom. The charge analysis suggests that the smallest species are doubly charged, and the large ones can accommodate 3-4 electrons. Enhanced adsorption on metal-supported ultra-thin film is not limited to metal species.
Calculations assign that Au
The phenomenon is more general in nature and present also for molecules with high electronic affinity such as O 2 and NO 2 . Understanding NO 2 adsorption characteristics is important, for example for NO x storage catalysts to control emissions from combustion in oxygen excess 124 .
On bulk-like MgO NO 2 adsorbs weakly and stays neutral 125
23
Relatively little attention has been paid on chemical reactions on MgO due to its inertness compared to more reactive reducible oxides such as TiO 2 . However, as discussed above ultrathin films display higher chemical reactivity compared to bulk MgO binding species with high electron affinity more strongly. The question then arises: can thin films systems facilitate chemical reactions ? difference in kinetics of CO 2 production from CO+O 2 and CO+NO reactions indicate that the healing of oxygen vacancies is the rate-limiting step 21 . DFT calculations demonstrate that this is indeed the case because NO adsorbs N-down on an oxygen vacancy and CO reaction with the "dangling" oxygen is a highly activated process 21 .
The adsorption and activation of O 2 is known to be difficult on Au 131 , yet its pivotal role in oxidation reactions on an Au catalyst is undeniable. Therefore, it is natural that first reactivity studies on ultra-thin film supported Au clusters include molecular oxygen. O 2 adsorption and activation on thin-film supported Au clusters has been theoretically analyzed on and is strongly linked to bridge the materials and pressure gaps to discover potential morphology changes of the catalyst particle under ambient conditions. The Au 14 cluster has an idealized C 2v symmetry with ten atoms at the edge and four inner atoms forming a rhombus-like structure. All the favourable adsorption sites reside at the edge of the cluster as the adsorption on the top of the cluster is endothermic. For all the studied structures adsorption energies range from from -1.15 eV for a single molecule to -0.6 eV for the 10'th molecule. The best adsorption structures for 1, 2, 6, and 10 oxygen molecules are given in relative to the orientation as shown in Table I showed evidence of facile hydrolysis [155][156][157] , where the active site was proposed to be at the metal-oxide interface. This was further supported by the DFT study of the system 158 . Interestingly, calculations show that isolated H-bonded water dimers facilitate barrierless water dissociation on a bare and Ag-supported MgO(100) surface 161 . Dissociation is further stabilised on MgO ultra-thin films, which is attributed to two factors. The presence of metal support stabilizes the final state of charged fragments owing to structural relaxations in oxide that induce a strong corrugation of the film. Furthermore, based on the analysis of interaction energies the ion-ion interaction is suggested to change to a dipole-dipole type interaction ensuing a weakly bound ion pair 161 . At monolayer water coverages different structural motifs corresponding to surface cells with different sizes have been employed in calculations 148 .
An example applied structural model is given in Figure 18 The properties of metal supported ultra-thin oxide films can be tuned without varying the film thickness via engineering the interface between the metal support and the oxide. This can be achieved by introducing impurity atoms into the metal oxide interface and by varying the species, which in turn has been predicted to modify the electronic characteristics of the system and can be seen as changes in the work function [163][164][165] . Also, interfacial vacancies can be induced in the supported oxide films. In these systems kinetic effects have been adsorption and dissociation on Ag-supported MgO. In the latter two cases, an Ag atom in the topmost Ag layer is replaced with either an impurity atom or a dopant atom. The oxidemetal interface modifications change neither the reaction mechanism, nor the adsorption geometries of the water molecule, nor the dissociation products OH − and H + ; they remain similar to those on the ideal MgO(100)/Ag(100) surface discussed in chapter IV C. Instead, any kind of the interface defects introduce changes in energetics: smaller for the adsorption energy of water and larger for the transition and final states. The transition state strongly resembles the final state and actually describing it as a transition state is a controversial issue since the H-OH bond is clearly already broken at this state. Comparison of the influence of an interfacial oxygen vacancy and O/Mg impurity atoms shows that the vacancy introduces the largest stabilization effect: the barrier height drops 42 % compared to the ideal MgO/Ag system. Actually, the dissociation does not require external thermal energy atall to proceed, which means that even a small number of hidden vacancies is sufficient to enhance water dissociation substantially. Among the studied 3d TM dopants, the Ti atom is the most reactive one, and it stabilises the initial, transition and final states by 0. 17 including the plane average density analysis, the electronic localization function, and the Bader charge analysis, unravel that while one vacancy electron stays in a support metal atom under the vacancy, the second electron goes to a Au atom; thus the vacancy is more like a F ++ center. The simple thermodynamic analysis -with charge transfer and electrostatic terms-renders the formation of F ++ thermodynamically possible. According to the calculations, the electron transfer process is independent of the film thickness (at the range of 1-3 layers), and takes place similarly on both Ag and Mo support metals.
The fact that Au binding energy to the surface vacancy is always more exothermic than to the site with a buried vacancy, leads to the thermodynamic driving force to extract the vacancy from the metal-oxide interface to the surface. Despite thermodynamic preference, the vacancy extraction from the interface to the oxide surface can be hindered by kinetic factors, that is, too high diffusion barriers. The diffusion pathway of an oxygen vacancy consists of jumps between two nearest neighbour sites. In bulk MgO this has been reported 41 to be highly demanding with activation energy of ∼ 4.25 eV 170 . On two or three layers thick, unsupported MgO slabs the barrier is markedly reduced being still, however, ∼ 2. Clearly, adsorption becomes weaker with the increasing distance between the Au and the dopant. The term ∆E coul demonstrates approximately 1/r behaviour, where r is the Audopant distance proving further support that energy change ∆E coul describes electrostatic interaction. Owing to strong screening by the polarized oxide, the Coulomb interaction is significant only at very short distances. In experiments the capping region is typically a few undoped CaO layers 37 . If the Mo is in the third layer the Au-Mo distance is approximately 8Å which corresponds the contribution of only ∼-0.5 eV while the adsorption energy is -3.0 eV, thus the electrostatic contribution to Au adsorption is minor on this oxide. The adsorption energy is dominated by the other two terms. The iono-covalent term is independent of a dopant as it describes the interaction between Au and ions in the oxide matrix but depends on the adsorption site. The ∆E redox term represents energy gain related to the electron exchange between Au and the dopant and corresponds to the dominant energy contribution in adsorption energy. It can be divided into two cases. Since the Au gains charge already upon the adsorption on a Ca-site of pristine CaO, the redox energy change is assigned to the system stabilization due to Mo oxidation as the more cationic dopant binds stronger to anionic O 2− . This is manifested as decreased distance between the dopant and the surrounding oxygens. Since this is a local effect, it does not depend on the depth of the dopant as seen in Figure 23. On an O-top site Au is initially neutral and thus the redox process includes charge transfer and stabilization due to structural relaxations around the dopant; again, the redox energy saturates quickly to a constant value. The redox energies differ by 1 eV such that the less exothermic value is found for the O-top site. This is due to Coulomb repulsion between negatively charged Au and negatively charged O, manifested also in a Au-O bond length, which increases ∼ 0.6Å upon Au charging. Figures 23 B) 34 . The reaction has not been thoroughly understood yet and for example, an active site for a homolytic CH 3 -H bond breaking step has remained elusive.
In the case of undervalent doped systems, the Born-Haber cycle analysis can also be applied to unravel the relevance of different energy terms for e.g., Au adsorption, which has been experimentally studied 39
49
Over the past years significant advances have been made in the preparation and experimental studies of tailored oxides as well as in the modeling their properties. These modified oxides have emerged as materials with unique properties, not encountered in their bulk counterparts. The selected examples discussed highlight the potential of simple oxides. They can be made chemically active and can operate as active catalysts or act as supports for catalyst particles. Two different schemes to modify the properties of simple oxides are addressed, namely growing oxides as ultra-thin films over metal supports and doping them with metal impurities. The discussion mainly focuses on theoretical studies for metal supported MgO films and in doped oxides on CaO and MgO systems, while other oxides are occasionally brought in attention for comparison. All the reviewed systems underline the potential and importance of density functional theory calculations and their role to unravel experimentally seen features but even more importantly to predict material properties, especially activity.
Despite the fact that computational methods have their limitations, for certain cases, such as buried vacancies, they are almost irreplaceable in addressing the characteristics and impact of hidden defects on adsorption.
The key adsorbates elaborated include electronegative species such as Au atoms and clusters, molecular oxygen, and NO 2 . The adsorption and dissociation of molecular oxygen is considered both on a bare metal-supported ultra-thin MgO and on an Au cluster over Ag-supported MgO. Interestingly, the calculations predict that the presence of the support metal facilitates simultaneous activation of several oxygen molecules. The adsorption and dissociation of non-electronegative water on ideal MgO/Ag is discussed both at low and high water coverages. In the latter case, the formation of strong hydrogen-hydrogen bonds leads to dissociation of a fraction of adsorbed water molecules. In addition, the modifications at the oxide-support interface are addressed in the two cases; on Au adsorption on a MgO/Ag(Mo) surface with an interfacial oxygen vacancy and water adsorption on MgO/Ag with an interfacial dopant. The role of thin-film systems is not limited to model systems but they have a variety of applications including catalysis, solid oxide fuel cells, gas sensors, corrosion protection, and biocompatible materials, just to mention a few. The spectrum and composition of oxide materials developed for different application is broad; some of these oxides are simple while there is an increasing number of fairly complex oxides. In the 50 future, we need to compute the electronic structure of these complex oxides reliably, which potentially have a variety of different defects, and simultaneously get a description for weak van der Waals interactions. In particular, calculation of line defects increases the system size, which in turn poses challenges for calculations.
On doped CaO and MgO oxides, the adsorption of an Au atom and an O 2 molecule is discussed. While the film thickness is an adjustable parameter in metal-supported thin film systems, in doped oxides the corresponding parameter is the dopant concentration.
Moreover, unlike the metal support, dopants can be mobile and they are also predicted to introduce vacancies into oxide. However, calculations indicate that the impact of over-valent dopants on electronegative adsorbates such as Au atoms and O 2 molecules, is similar to the metal support. Thus adsorption is enhanced, the species become negatively charged, and on both supports Au clusters favour planar geometries. With the help of the Born-Haber cycle the enhanced binding is attributed to energy gain owing to simultaneous electron transfer and lattice relaxations. In an idealized system, the ionization energy of the TM dopant explains the variation of Au adsorption energy from one dopant to the other but in reality dopant induced vacancies have an influence on electron transfer processes and modify this ideal rule. While in metal-supported ultra-thin oxide films systems one can gain better control over structural and electronic characteristics, the atomic structure of doped oxides displays larger plasticity and uncertainty. The model system studies of these materials make the basis to identify the key factors determining the function of the material, which paves the way for examination of less-defined oxides.
Doped oxides also have a variety of applications ranging from catalysis and photovoltaic to chemical sensing, solid oxide fuel cells, and coatings. In many cases the doped oxide is in the thin-film arrangement. The presence of dopants typically affects a number of point defects, such as oxygen vacancies, which then strongly impact on surface chemistry of an oxide. From the theoretical point of view, the concentration and distribution of dopants and vacancies substantially enhance the computational burden and add complexity. Although possible, special care must be taken when the Born-Haber cycle is applied to rationalize chemical reactions on complex, doped oxide surfaces. This is because sufficiently accurate description of the electronic structure is needed and because the impact of vacancies must be included to the analysis. Furthermore, the identification of an active site is usually more demanding than on metal surfaces and nanoparticles owing to a larger number of different 51 factors such as dopant and vacancy concentration, which might influence the nature and characteristics of the active site. One more thing that makes DFT calculations particularly challenging, is the lack of experimental methods, that could reveal the morphology and composition of oxides under reaction conditions. Therefore, there is an urgent need to develop computational methods and concepts to simulate oxide characteristics and establish their key descriptors to advance in one of the most challenging areas of Material Science. | 5,955.6 | 2014-12-01T00:00:00.000 | [
"Chemistry",
"Physics"
] |
Quantifying the Bullwhip Effect in a Reverse Supply Chain: The Impact of Different Forecasting Methods
The reason for this study is that the bullwhip effect can pose very serious consequences for enterprises, such as increased production costs, additional manufacturing costs, excessive inventory levels, excess storage costs, large capital overstocking, and excessive transportation costs. Thus, the problem for this study is that quantifying the bullwhip effect in a reverse supply chain and comparing the impact of different forecasting methods on it. The objective of this paper is the bullwhip effect (BE) in a reverse supply chain (RSC). In particular, this study proposes a quantitative expression of the BE in a RSC, that is, BE R (cid:31) Var ( q t ) / Var ( r t ) , and analyzes the impact of different forecasting methods (e.g., the moving average technique (MA), the exponential smoothing technique (ES), and the minimum mean square error forecasting technique (MMSE)) on the bullwhip effect. We evaluate the conditions under which the collector should select different forecasting methods based on the BE. We use simulation date and get some conclusions that, in some cases when using the MMSE method, the BE does not exist in a RSC. This finding is significantly different from the results on the BE in a forward supply chain. Moreover, the MMSE method can reduce the lead-time demand forecast error to the greatest possible extent, which allows the BE to reach the lowest level.
Introduction
With increasing awareness of environmental protection issues, many countries and rms pay greater attention to activities related to product recycling. Reverse logistics is an emerging research eld, and it involves the process of moving goods from their nal destination for the purpose of properly disposing of the goods or capturing value and pro ting from them. Many scholars have proposed a denition of reverse logistics. For example, reverse logistics describes the recycling and remanufacturing activities undertaken by an enterprise that collects used products from the consumer in supply chain management. is de nition implies that reverse logistics involves saving raw materials in the production process, recycling, reuse and recovery of used products, reusing packaging, and dismantling or repairing defective products. From a wider societal perspective, the implementation of reverse logistics can e ectively protect the environment, reduce the consumption of energy and nonrenewable resources, improve the utilization rate of resources, and realize the sustainable development strategy. For enterprises, e ective reverse logistics operations can not only reduce material costs and increase an enterprise's income, but also enhance their corporate image by improving customer satisfaction, increase the information exchange at each node of the supply chain (SC), improve market share, and establish a competitive advantage.
us, the reverse logistics networks, structure performance, and operations modeling have become hot issues in academic research.
e bullwhip e ect (BE) is a phenomenon that is observed when there is an ampli cation of demand uctuation up the supply. As demand information becomes distorted, an enlargement can occur in the process of the transmission of demand information from downstream to upstream. Moreover, the uctuation in demand in upstream enterprises is greater than that observed in downstream enterprises, which can cause the bullwhip e ect (BE). e BE is the most important performance indicator in the SC structure, and it is also the most important performance index in the SC operation. e BE can pose extremely serious consequences for enterprises, such as increased production costs, additional manufacturing costs, excessive inventory levels, excess storage costs, large capital overstocking, and excessive transportation costs. A substantial body of literature has examined the BE. Forrester [1] first identified the existence of the BE. ereafter, many scholars studied the existence of the BE, aiming to identify its causes and assessing how it may be reduced. Although many scholars have carried out a lot of research to examine the strategic, tactical, and operational levels of reverse logistics, few scholars discussed the quantification of the BE and its impact on reverse logistics. erefore, this paper develops a quantitative expression of the BE in a reverse supply chain (RSC). e impact of different forecasting techniques on the BE in an RSC is also compared. Finally, the relevant measures to reduce the BE in a RSC are proposed.
Of course, many firms have also found the bullwhip effect in their operational practices. For example, in the middle of 1990s, when workers at Procter & Gamble were examining order patterns for their best-selling baby diapers, they notice a strange phenomenon that the selling quantity of this product is fairly stable and does not fluctuate much, but when examining orders in the distribution center, we were surprised to find that the fluctuation increased significantly. e distribution center said that it placed orders based on aggregate demand for orders from vendors. Further study found that retailers are often based on historical sales volume and real sales forecast and determine the quantity of a more objective, but in order to guarantee that the quantity of goods is available on time, as well as the ability to adapt to incremental changes in customer requirements, they will usually request that the forecast order quantity must be enlarged to wholesalers. As a result, consumer demand has been amplified. e main purpose of this paper is that we find and derive the mathematical expression of bullwhip effect in reverse supply chain in theoretically. is is the biggest difference comparing the previous studies. e mathematical expression of the bullwhip effect that we derive does not change depending on the forecasting techniques used. Based on this, we analyze the different influence of different forecasting techniques on bullwhip effect in reverse supply chain. In other words, the main contributions of this paper are as follows. First, we come up with the quantitative expression of the BE in a RSC by using different forecasting techniques (i.e., the moving average technique (MA), the exponential smoothing technique (ES), and the minimum mean square error forecasting technique (MMSE)). Second, we compare the influence of different forecasting techniques on the BE in a RSC. ird, this paper proposes relevant measures that can be used to reduce the BE in a RSC, and we can obtain some important managerial insights. e remainder of this paper is organized as follows. Section 2 reviews relevant literature. Section 3 presents a description of the problem and describes the modeling. Section 4 determines the BE using different forecasting methods, that is, the moving average technique (MA), the exponential smoothing technique (ES), and the minimum mean square error forecasting technique (MMSE). e simulation and results of the analysis are introduced in Section 5. e final section presents the conclusion and direction for future research.
Literature Review
In traditional forward SCs, the BE refers to the phenomenon of amplification of demand variability from the point of final demand to the point of origin [2]. is phenomenon can lead to substantial problems that affect SC performance, such as superfluous inventory, erroneous product forecasts, and high costs for correction [3]. e BE has become one of the main obstacles affecting the efficiency of the SC. us, it has attracted the attention of numerous administrators and scholars.
Early studies mainly focused on the existence of the BE and recognized its causes in traditional forward SCs (Forrester [1]; Lee et al. [2]; Lee et al. [3]). Lee et al. [4] discussed the BE in a two-level SC under an AR(1) demand process. Alwan et al. [5] studied the BE under the ARMA(1,1) demand process. In addition, the ARMA(p,q) demand process was discussed by Gaalman and Disney [6] and Gaalman et al. [7]. Michael [8] analyzed the bullwhip effect problem from the carriers' viewpoint under the AR(1) demand process. Moreover, many scholars discussed the BE in a two-stage forward SC using control approaches. Holt et al. [9] developed the HMMS control model in a two-stage forward SC and pointed out that this model could effectively balance the relationship between ordering from the retailer and ordering from the supplier. Blinder et al. [10] proposed that it could reduce the BE by using the (S,S) ordering strategy. Jose and Rafael [11] used a web-based supply chain simulator to demonstrate the potential benefits of using Electronic Data Interchange (EDI) in supply chain management. Moreover, they pointed out that it could reduce the bullwhip effect by using EDI technique. Baganha et al. [12] designed a particular inventory control strategy and highlighted that it could reduce the fluctuation of demand information. In addition, more scholars studied the BE using the discrete control theory, H ∞ control theory, control-based forecasting technique, and O-S feedback control method in twostage forward SCs (e.g., Towill [13]; Huang et al. [14]; Disney et al. [15]; Rong et al. [16]).
With economic and societal development, more and more companies are involved in the SC, and many scholars attempted to analyze the BE in a multistage forward SC using different approaches. Li et al. [17] derived closed-form formula to analytically describe how the bullwhip and antibullwhip effects originated initially and then evolved over time and spaced in the supply chain. Vicente et al. [18] measured the bullwhip effect through four metrics: the echelon average inventory, the echelon inventory variance ratio, the echelon average order, and the echelon order rate variance ratio. Zhu et al. [19] investigated the factors that impacted the bullwhip effect in the oil and gas supply chain using case study evidence from six companies in North America, which cover refining and marketing, exploration and production, integrated oil and gas, and drilling. Yao et al. [20] estimated the product level bullwhip effect using various methods, analyzed consequences of its different measurements and aggregations, and examined its impact on supply chain performance in terms of inventory ratio and stock outs. Yin [21] have considered the market competition among retailers and measured the bullwhip effect, in which multiple retailers exhibit AR(1) demand processes, and the degree of market competition was captured with copula. Lu et al. [22] have adopted regression models to test the proposed model and conducted a series of robustness tests by using moving average forecasting methods. Costantino et al. [22] studied the influence of demand sharing on order strategy using the simulation method. Hossein et al. [23] quantified the BE, order rate variance ratio (OVR), and inventory variance ratio (IV) in a three-stage forward SC with multiple retailers. Marieh et al. [24] investigated the measure of the BE in a three-stage forward SC. Alexandre et al. [25]performed a simulation-based study to investigate the interrelations of the structural and operational dynamics in the forward SC. Moreover, they pointed out that the forward SC managers need to take into account the risk of bullwhip effect during the capacity disruption and recovery periods. Ki and Jae [26] built a four-echelon supply chain simulation model where each echelon shares some of the customer demand forecast information with a retailer, the lowest echelon. What is more, they analyzed the impact of information sharing on the bullwhip effect. Bray et al. [27] modeled a single-supplier, 73-store forward supply chain as a dynamic discrete choice problem. ey estimate the model with transaction-level data, spanning 3251 products and 1370 days. Ojha et al. [28] used simulation to investigate the impact of information sharing on both the bullwhip effect (BWE) and the order-fulfillment performance (OFP) in multiechelon forward supply chain system. Li [29] researched the bullwhip effect in a two-echelon forward supply chain consisting of one single supplier and multiple retailers, and the vertical and horizontal cooperation game for carbon emission reduction was analyzed under carbon tax scheme. Yao et al. [20] estimated the product level bullwhip effect using various methods, analyzed consequences of its different measurements and aggregations, and examined its impact on supply chain performance in terms of inventory ratio and stockouts. Yin [21] measured the bullwhip effect in a two-stage supply chain with one supplier and multiple retailers, in which multiple retailers exhibit AR(1) demand processes, and the degree of market competition was captured with copula. e simple two-stage forward SC modeling assumption has been widely used to study the BE. However, due to the complexity of the forward SC network's structure, the simple two-stage forward SC modeling assumption is outdated. Recently, some scholars discussed the BE in a two-stage forward SC network (Zhang and Zhao [30]; Zhang and Yuan [31]; Yuan and Zhu [32]). Nonetheless, most of these studies assumed that customer demand follows the AR(1) autoregressive process, and they mainly discussed SC coordination, and the BE could be reduced. Yuan and Zhu [32] provided three quantitative models of the BE in a two-stage forward SC network.
It is known that a great number of scholars studied the BE in two-stage or multistage forward SCs. On the contrary, fewer scholars discussed the reverse BE in a forward SC. e term "reverse bullwhip effect" (RBE) was first proposed by Svensson [33], who analyzed the RBE in intraorganizational echelons. He pointed out that the RBE occurs when there is a high degree of postponement in inbound logistics flows, and a high degree of speculation in outbound logistics flows. Ozelkan et al. [34] analyzed the impact of procurement price variability from upstream to downstream in a forward SC. Ozelkan and Cakanyildirim [35] investigated the conditions that lead to the amplification of price variations moving from upstream to downstream in a SC, which is referred to as the "reverse bullwhip effect in pricing" (RBEP). Ozelkan et al. [36] investigated the RBEP conditions for SCs, in which joint replenishment and pricing decisions are made.
e main difference between this paper and the existing literature lies in the following aspects: (1) as discussed in the literature review above, most papers mainly analyze the bullwhip effect in forward supply chain including two-level or multilevel supply chain. However, we give the quantitative expression of the bullwhip effect in reverse supply chain. (2) Most of the above papers discuss the impact of three different forecasting techniques on the bullwhip effect in forward supply chain; on the contrary, in this paper, we discuss the impact of different forecasting techniques on the bullwhip effect in reverse supply chain and propose relevant measures that can be used to reduce the bullwhip in reverse supply chain.
Problem Description.
We consider a two-stage RSC consisting of one collector and one remanufacturer (see Figure 1). e collector is the only supplier of used products to the remanufacturer. e trading activities occur over an infinitely discrete period t, where t ∈ (− ∞, 0, +∞). At the end of period t, according to the past data on the supply of used products, the collector estimates the quantity of used products in period t + 1 by using different forecasting methods. Moreover, according to a certain inventory strategy, the collector determines the quantity of used products that should be transferred to the remanufacturer. e collector's lead time is a constant that can be expressed as l, and the remanufacturer can receive the used products at the beginning of the period t + l + 1.
Description of Parameters.
e relevant symbols that will be used throughout the paper are described and explained as follows: (i) r t : the supply quantity of used products from customers to the collector; (ii) q t : the quantity of used products transferred from the collector to the remanufacturer at the end of period t; (iii) ρ: autocorrelation coefficient; (iv) μ: nonnegative constant; (v) ε t : independent identically distributed random variable; (vi) S t : the collector's highest inventory quantity at the end of period t; (vii) S t− 1 : the collector's highest inventory quantity at the end of period t − 1; (viii) r l t : the estimate of the leading-time supply quantity using different forecasting methods; (ix) z: the constant to achieve a desired service level; (x) δ l t : the estimate of the standard deviation of the l period forecasting error of collector; (xi) p: the moving average period; (xii) l: the collector's leading time; (xiii) α: the collector's smoothing constant.
Modeling Assumption.
In order to make the quantitative expression of the BE more meaningful in practical terms in the case of the RSC, we made the following assumptions: (1) e collector only collects one kind of used products; (2) e quantity of used products from customers to the collector r t belongs to an AR(1) autocorrelation process; (3) We only focus on the quantitative expression of the BE for the collector. us, we assume that the collector can forecast the supply of used products from customers but will not share the information with the remanufacturer. (4) Assume that the quantity of used products from customers to the collector r t belongs to an AR(1) autocorrelation process: where μ is a nonnegative constant, ρ is the autocorrelation coefficient |ρ| < 1, and ε t is an independent identically distributed random variable with a zero mean and variance δ 2 . us, for any period t, it can be written as (2) (5) Assume that the quantity of used products transferred from the collector to the remanufacturer is q t , which can be expressed relatively to the supply quantity of used products from customers to the collector r t as where S t is the collector's highest inventory quantity at the end of period t, and S t− 1 is the collector's highest inventory quantity at the end of period t − 1, which is estimated from the observed the supply quantity of used products from customers to the collector as equation (4).
where r l t are estimates of the lead-time supply quantity using different forecasting methods, z is a constant to achieve a desired service level, and δ l t is the estimate of the standard deviation of the l period forecasting error of the recycler.
us, the quantity of used products transferred from the collector to the remanufacturer q t can be calculated relative to the estimates of the lead-time supply quantity r l t as Before deriving the quantitative expression of the BE in a RSC, it is necessary to define it first. It is widely accepted that, in a traditional forward SC, the BE is defined as the phenomenon of amplification of demand variability from downstream to upstream (Lee et al. [2]). Similarly, in a RSC, the BE refers to the amplification of supply variability of used products as one moves upstream in the RSC, that is, from customers to the remanufacturer. It is expressed as BE R � Var(q t )/Var(r t ). is means that the amplification in the quantity of used products transferred from the collector to the remanufacturer is greater than that of the supply quantity of used products from customers to the collector. (5), we can determine the mean lead-time supply quantity r l t using the MA method:
The BE in a RSC Using Different Forecasting Methods
where p is the moving average period. en, the quantity of used products transferred from the collector to the remanufacturer q t in equation (5) can be formulated as follows: Lemma 1. Using the MA forecasting method, the estimate of the standard deviation of the l period forecast error of collector δ l t is a constant and can be expressed as follows: Proof: Using the MA forecasting method, the estimate of the standard deviation of the l period forecast error of collector δ l t can be expressed as follows: where r t � μ + ρr t− 1 + ε t . en, the quantity of used products transferred from the collector to the remanufacturer q t in (7) can be formulated as follows: us, the variance of the quantity of used products q t in equation (10) can be derived as follows: It needs to be pointed out that, in equation (11), it can be proved that Cov r t− 1 , r t � ρVar r t , Cov r t− 1 , r t− p− 1 � ρ p Var r t , Cov r t , r t− p− 1 � ρ p+1 Var r t . (12) Theorem 1. In two-level reverse SCs, when the collector estimates the supply quantity of used products by using MA forecasting technique, the quantitative expression of the BE is as follows: where Λ � l/p.
e BE in a RSC Using the ES Technique.
Similarly, based on equation (5), we can determine the mean lead-time supply quantity r l t using the ES technique: where α (0 < α < 1) is a smoothing constant for the collector. en, the transfer quantity of the used products q t in equation (5) can be formulated as follows:
Mathematical Problems in Engineering
(15) Lemma 2. Using the ES forecasting technique, the estimate of the standard deviation of the l period forecast error of collector δ l t is a constant and can be expressed as follows: Proof: Using the ES forecasting technique, the estimate of the standard deviation of the l period forecast error of recycler δ l t can be expressed as follows: where Var(r l t ) � Var(l 2 (μ + ρr t− 1 + ε t )) � l 2 δ 2 + δ 2 /1 − ρ 2 .
Var r l t � l 2 Var r t en, the supply quantity of the waste product q t in equation (15) can be formulated as follows: us, the variance of the transfer quantity of used products q t in equation (19) can be derived as follows: It needs to be pointed out that, in equation (20), it is easy to prove that 6 Mathematical Problems in Engineering Var r t , Var r t , Theorem 2. In two-stage reverse SCs, when the collector estimates the supply quantity of used products by using the ES forecasting technique, the quantitative expression of the BE is as follows: where Ω � α/1 − (1 − α)ρ.
e BE in a RSC Using the MMSE Method. Box and
Jenkins have pointed out that the demand forecasts value d t+i is the conditional expectation of historical demand information for the period t + i(i � 0, 1, 2, . . .), i.e., Particularly for the AR(1) process, d t+i � E(d t+i |d t− 1 ). In this paper, the supply quantity of used products from the customer to collector at the end of period t follows the AR(1) process, such that en, the transfer quantity of used product q t in equation (5) can be formulated as follows: Proof: Using the MMSE forecasting method, the estimate of the standard deviation of the l period forecast error of collector δ l t can be expressed as follows: where r t � μ + ρr t− 1 + ε t . en, the transfer quantity of used products q t in equation (24) can be formulated as follows: us, the variance of the transfer quantity of used products q t in equation (19) can be derived as follows: Mathematical Problems in Engineering It needs to be pointed out that, in equation (28), it is easy to prove that
Mathematical Problems in Engineering
Theorem 3. In two-stage reverse SCs, when the collector estimates the supply quantity of used product by using MMSE forecasting method, the quantitative expression of the BE is as follows: where Θ � μ/1 − ρ.
e above results are summarized in Table 1.
Numerical Examples and Results
As can be seen from the quantitative expression of the BE in a RSC using three types of forecasting methods, the main factors affecting the BE include the autocorrelation coefficient, the moving average period, the collector's lead-time, and the smoothing constant for the collector. In order to analyze the influence of different factors on the BE, the simulation analysis is divided into two steps. First, we analyze the influence of the relevant factors on the BE when using different forecasting methods. We then compare the influence of different forecasting methods on the BE. Figure 2 illustrates that the BE in an RSC decreases quickly as the autocorrelation coefficient ρ varies from − 1 to − 0.4. But when ρ is greater than − 0.4, the BE shows a steady state. Additionally, the BE increases with the increase of leading time l. As shown in Figure 3, we can see that the influence of ρ on the BE when the moving average period p takes different values, which is similar to that when l takes different values. However, as p increases, the BE will decrease. at is, it is possible to reduce the BE in an RSC by increasing the moving average period, which is consistent with the result in a forward SC (Lee et al. [2]).
As can be seen in Figure 4, the BE in an RSC decreases as p varies from 1 to 4. But when p is greater than 4, the BE shows a steady state. Moreover, the BE decreases with the increase of ρ. As shown in Figure 5, the BE in a RSC decreases quickly when p is less than 4. After that, the BE tends to be stable, which is also consistent with the result in a forward SC (see Lee et al. [2]).
As shown in Figures 6 and 7, when l is less than 4, the BE is relatively stable. But when l is greater than 4, the BE increases quickly. In other words, l has a positive effect on the BE in an RSC. Figures 8-13 illustrate the influence of relevant factors on the BE using the ES method. Figures 8 and 9 show that the BE in an RSC decreases as α varies from 0.1 to 0.4. When the α value varies from 0.4 to 0.6, the BE is relatively stable. However, when the α value is greater than 0.6, the BE increases quickly. In contrast, l has a positive effect on the BE, while a smaller BE is associated with a greater ρ. It is worth noting that the smoothing constant α and the lead time l have similar effects on the BE in a forward SC (Alwan et al. [5]). Figures 10 and 11 also illustrate that ρ has a negative effect on the BE in a RSC, while Figures 12 and 13 also illustrate that the lead time l has a positive effect on the BE. We can see that when using the ES method, the lead time l and the autocorrelation coefficient ρ also have opposite effects on the BE. Figures 14 and 15 depict the influence of relevant factors on the BE using the MMSE method. As shown in Figure 14, when ρ is relatively small, the BE decreases sharply as l varies from 0 to 1, but it increases quickly as l is greater than 1. However, when ρ is large, the BE first increases slowly as l varies from 0 to 3, and after that, it increases rapidly. is is different from the results obtained using the previous two forecasting methods. As can be seen in Figure 15, the BE in a RSC is relatively stable when the ρ value varies from -1 to 0.2. After that, the BE increases quickly.
e Influence of Relevant Factors on the BE Using the MMSE Method.
is differs from the results obtained using the previous two forecasting methods. Furthermore, in some cases, when ρ varies from -1 to 0.2, the BE does not exist in a RSC, which is significantly different from the results on the BE in a forward SC.
Comparing the Influence of Different Forecasting
Methods on the Bullwhip Effect. From the above analysis, it is evident that the autocorrelation coefficient ρ and the lead time l have an important influence on the BE under three kinds of forecasting methods. Next, we will compare the influence of different forecasting techniques on the BE. Assume that α � 2/(p + 1), p � 4, α � 0.4. Figure 16 illustrates the impact of ρ and l on the BE in a RSC under three kinds of forecasting techniques.
In Figure 16, we can see that the BE in a RSC reaches the largest level under the ES method, which is comparable to that in a forward SC. In addition, when the autocorrelation coefficient is less than 0.6, the BE is at the lowest level under the MMSE method; when the autocorrelation coefficient is greater than 0.6, the BE is at the lowest level. To sum up, we can obtain the following managerial insights: (i) Managerial insight 1: when the collector predicts the supply of used products from customers using the MA method, both the autocorrelation coefficient ρ and the moving average period p have a negative effect on the BE, while the lead time l has a positive effect on the BE. us, the collector could reduce the BE by increasing the autocorrelation coefficient and the moving average period, or by reducing the lead time. (ii) Managerial insight 2: when the collector predicts the supply of used products from customers using the ES method, the BE is at the lowest level when the smoothing constant α varies from 0.4 to 0.6. us, the collector should choose an appropriate smoothing constant to reduce the BE. Managerial insight 3: when the autocorrelation coefficient and the lead time satisfy eorem 4, the collector should use the MMSE method to predict the supply of used products from customers so as to reduce the BE. Otherwise, the collector should use the MA method.
Conclusion
In this paper, we examine the influence of different forecasting methods on the BE in a two-stage RSC that consists of one collector and one remanufacturer. First, we develop a quantitative expression of the BE in a RSC. We then analyze the influence of the autocorrelation coefficient, the leadtime, the moving average period, and other factors on the BE in a RSC. Finally, we analyze the conditions under which using different forecasting methods can reduce the BE. We reached the following conclusions: (1) When the collector predicts the supply of used products from customers using the MA method, both the autocorrelation coefficient ρ and the moving average period p have a negative effect on the BE, while the lead-time l has a positive effect on the BE. (2) In some cases, when ρ varies from -1 to 0.2 and using the MMSE method, the BE does not exist in a RSC, which is significantly different from the result on the BE in a forward SC. (3) To reduce the BE in a RSC, when the autocorrelation coefficient and the lead-time satisfy eorem 4, the collector should use the MMSE method to predict the supply of used products from customers. Otherwise, the collector should use the MA method.
In this paper, we only study the impact of different forecasting methods on the BE in a two-stage RSC. In the future, we could discuss the BE in two competitive RSCs, each of which consists of one collector and one remanufacturer.
Data Availability
e data used to support the finds of this study are available from the corresponding author upon request.
Conflicts of Interest
e authors declare that they have no conflicts of interest.
Authors' Contributions
Y. X. G. and Z. X. Q. conceptualized the study and wrote the original study. W. M. and Z. D. L. reviewed and edited the article. | 7,315 | 2022-05-09T00:00:00.000 | [
"Business",
"Economics",
"Engineering"
] |
A Data Acquisition and Processing Method for Edge Computing Robotic Arm Behavior Recognition
Edge computing refers to the use of an open platform on the side close to the object or data source to integrate network, computing, storage, and core application functions to provide the latest nearby services. With the development of edge computing, the cost of data acquisition has been reduced, and the e ffi ciency has been improved. However, at present, there is no in-depth research on edge computing for robot arm behavior recognition. This paper aims to study the data acquisition and processing methods of robotic arm behavior recognition through edge computing technology. A gesture recognition method based on Cauchy distribution and grey correlation threshold is proposed, which improves the e ffi ciency of data processing and has great research signi fi cance. In edge computing, the use of Cauchy distribution processing is more impressive; compared with empirical distribution, the algorithm optimization can reach at least 10%. Experiments show that the static gesture recognition method used in this paper is simple and high in recognition and has good robustness and the accuracy rate can basically reach more than 90%. In the case
Introduction
The society is gradually entering the era of "big data", and with the advent of cloud computing, its ability to operate and use big data collection is also increasing. Cloud computing is actively supported by advantages including low operating costs, low weight, and ease of use and maintenance. The cloud computing industry is actively developing in China. Edge computing was initially the primary technology to address 5G latency, but has recently been introduced into new areas such as IoT and the Internet. The problem of cloud computing can also be solved with the help of modern information technology. The intelligent services provided by advanced computer technology meet key requirements such as flexible connection, real-time operation, data upgrade, software intelligence and security, and data basic processing. Robotic arms are the most widely used robots. It exists in the construction, industrial, healthcare, entertainment, military, electronics, manufacturing, and space exploration industries. Robotic arms come in many different forms, but they all share common characteristics. The ability to select a course of action and precisely point-and-shoot in threedimensional (or two-dimensional) space. The traditional mechanical arm can only perform simple repetitive actions such as translation and grabbing, but when the mechanical arm is combined with the information system, it can complete complex operations such as cooking and wine mixing. At this time, it is necessary to identify and analyze the complex motion behavior of the manipulator. However, the traditional data acquisition and processing methods cannot analyze the data quickly, which leads to low efficiency. Therefore, it is necessary to collect and process the data of manipulator behavior recognition based on edge calculation.
Aiming at the data acquisition and processing of the manipulator, this paper has made great innovations in data acquisition and uploading. We first select the optimal threshold value, fully considering the accuracy of the recommendation achieved by the threshold value and the timeconsuming situation and also consider the role of the threshold value in the subsequent selection of the number of neighbors. The optimal threshold is selected to realize the clustering of gray projects, which makes the internal common features of the projects prominent, and the external differences of the projects are relatively large. Calculating the similarity of items on the obtained clusters, and according to the fixed number of neighbors, sort the calculated similarities from large to small in turn and select a fixed number of neighbors. The scoring prediction and recommendation of the item are completed, and the performance of the recommendation algorithm is verified experimentally. We have done in-depth research on static gesture recognition based on multiparameter features from image acquisition to gesture recognition. The RGB-D image information is obtained by using the Kinect depth sensor, and then the gesture image is segmented by the adaptive depth threshold fused with skin color, and the corresponding preprocessing is performed to extract the outline of the gesture.
Related Work
With the emergence and popularity of Internet of Things (IoT) cloud services, edge computing, a new technology that requires data processing over the Internet, has emerged. Chakraborty proposed a local pattern descriptor in higherorder derivative space for face recognition. His proposed descriptor significantly reduced the extraction and matching time, while the recognition rate of the descriptor is almost comparable to the existing state-of-the-art methods [1]. Taleb's paper presented a survey of mobile edge computing (MEC), outlined current standardization activities, and further elaborated on the challenges of open research [2]. Jiang proposed a robust trajectory tracking control method for robotic arms based on H control theory, and the field operation test further verified the engineering practicability of the control method in macro and micro aspects [3]. For the rapid picking of lilies, Jiang designed a mechanical arm picking structure scheme, using a system consisting of an end effector, a manipulator, and a control system. The kinematics and picking experiments of the robotic arm were carried out in the experimental field in the natural environment through the robotic physics machine platform. The results show that the position error of the robot arm from the end of the arm is less than 12 mm, and the picking success rate is 83.33% [4]. The related research did not pay attention to saving the data of the manipulator, and even did not simulate the trajectory of the manipulator.
Maffezzoni's paper provided a realistic phase-domain modeling and simulation approach for oscillator arrays that is able to account for associated device non-idealities. This model was used to study the associative memory performance of an array consisting of resonant LC oscillators [5]. To practically implement brain-like computing in scalable physical systems, Kumar investigated a network of coupled micromachined oscillators. He performed numerical simulations of this all-pair fully coupled nonlinear oscillator array in the presence of randomness and demonstrated its ability to synchronize and store information with relative phase differences when synchronized [6]. Liu proposed an opensource face recognition method with deep representation called VIPLFaceNet, which is a 10-layer convolutional neural network consisting of 7 convolutional layers and 3 fully connected layers. Compared to the famous AlexNet, VIPL-FaceNet only needed 20% training time and 60% testing time, but combined with actual LFW facial recognition metrics, it achieved a 40% error reduction. Their VIPLFaceNet achieved an average accuracy of 98.60% on LFW using a single network [7]. Biyani presented a new software package called Focus, which provided the functionality needed to remotely monitor the progress of data collection and data processing, and the rapid detection of any errors that may occur greatly increased the productivity of electron microscopy recording sessions [8]. However, none of them solve practical problems such as the redundancy of edge computing in data processing, and the following content will conduct in-depth research on these problems.
Edge Computing System Design and Data
Sampling Method 3.1. Edge Computing System Design. The single-user scenario is shown in Figure 1. Whether it is cloud, fog, or edge computing, it is only a method or mode to realize the computing technology needed by the Internet of Things, intelligent manufacturing, etc. On the basis of a single system, a reserved sensor interface is used to complete the realization of the sensor acquisition function. As shown in Figure 2, by connecting the sensors that meet the requirements of the interface hardware with the edge device, the software is used to complete the preprocessing of the collected data, thereby forming the underlying data acquisition module. This can not only expand the scope of application of the sensor, but also ensure that the collected data is uploaded to the cloud in time, shortening the delay of data processing and feedback [9].
The architecture of edge computing includes four domains: equipment, network, data, and application. Platform providers mainly provide hardware and software infrastructure in network interconnection (including bus), computing power, data storage, and application. In addition to the collection of actual physical data, with the continuous development of informatization today, a large amount of data existing on the Internet is also worthy of attention and utilization. Therefore, for the data that can be accessed on the network, especially those web page data presented 2 Wireless Communications and Mobile Computing in HTML and other formats, people will use various methods to download the web page codes and the data behind them and process them. As shown in Figure 3, the system utilizes the characteristics of a large number of edge devices and low energy consumption and can establish a multi-threaded or long-term monitoring data capture system by providing it with a crawler program. For example, users can carry sensors through the edge device network to monitor the air quality of a certain area and at the same time use web crawler technology to query the indexes of various local social activities, such as factory operating conditions and regional weather. In this way, it is compared with the collected air parameters to analyze the factors that affect the local air quality. The rapid development of global smart phones has promoted the development of mobile terminals and "edge computing." The intelligent society with the internet of things and the perception of everything is accompanied by the development of the Internet of Things, and the edge computing system has emerged accordingly. In the edge computing system, the platform needs to be built by the cloud computing center and edge devices. The advantage of this architecture is that when the bottom layer needs to call the data analysis function, the edge device can complete the timely feedback, shorten the response time, and ensure the real-time response characteristics of edge computing [10]. For the Internet of Things, breakthroughs in edge computing technology mean that many controls will be implemented through local devices without being handed over to the cloud, and the processing will be completed at the local edge computing layer. This will undoubtedly greatly improve processing efficiency and reduce the load on the cloud. Due to being closer to the user, it can also provide users with a faster response and solve their needs at the edge.
Host Component Analysis
Method. Principal component analysis is a statistical method. Through orthogonal transformation, a group of variables that may be related are converted into a group of linearly unrelated variables, and the converted variables are called principal components. The traditional principal component analysis algorithm (PCA) is a method to minimize measurement features and is an important tool for measurement analysis. It has a predictable learning path [11]. Now PCA is continued to be derived for the mean least squares error of the object.
The sample matrix after zero mean processing is written as Y ′ , and Y ′ = fy 1 ′ , y 2 ′ , ⋯, y n ′ g. After the sample zero-mean centering, its autocorrelation function can be expressed as Let the optimal transformation vector be Cd ∈ R l , and find the eigenvalues and eigenvectors of B y : The largest eigenvector obtained from this characteristic formula is the optimal PCA transform vector [12]. In the PCA dimensionality reduction process, it is usually necessary to manually adjust the reduced size [13]. The transformation matrix after dimensionality reduction can be 3.2.2. Kernel Principal Component Analysis. If the purpose of PCA is to attenuate nonlinear correlations between a given data set ðY = fy 1, y 1, ⋯ y j, ⋯ , y n g, y i ∈ R l Þ, their covariances can be represented on a linear feature space F instead of the nonlinear representation in the original input space [14].
It is assumed that the mapping of the feature vector in the high-dimensional space is also zero-mean, that is, ∑ n k=1 Φðy k Þ = 0 and Φð•Þ are nonlinear mapping functions that map the input feature vector to the high-dimensional space F. Then, diagonalize the covariance matrix: Next, the eigenvalues and eigenvectors are solved by solving the characteristic formula. The specific content of B F ν can be deduced according to Formula (6): In the formula, hu, vi represents the inner product between u and v. The formula shows that all eigenvectors v of μ ≠ 0 are composed of Φðy 1 Þ, ⋯, Φðy n Þ, so we can use v = ∑ n j=1 to represent the eigenvector, and Formula (7) can also be written as Combining (8) and (9), we can get Now, a matrix K of size n × n is assumed, where K ij = hΦðy i Þ, Φðy j Þi; then, the left-hand side of Formula (10) can be written as The right-hand side of the formula can be written as Combining Formulas (11) and (12), we get
Cauchy Distribution Function Data Sampling and
Processing. The distribution is estimated from the sample using the Cauchy distribution function, mainly the estimation of the parameters of the Cauchy distribution, and then the inverse cumulative distribution function is used for sampling [15]. A description of the Cauchy distribution function and parameter estimation and sampling methods is first given [16]. The Cauchy distribution, also known as the Cauchy-Lorentz distribution, is a continuous distribution function. If the probability density function of random variable ξ is Among them, γ, y 0 are constants, and γ > 0, then ξ is said to obey the Cauchy distribution with parameters γ, y 0 and denoted as ξ~Cðγ, y 0 Þ. The cumulative distribution function of ξ~Cðγ, y 0 Þ is In the formula, y 0 is the position parameter that defines the location of the peak of the distribution, and γ is the scale parameter that is half width at half the maximum value. And F ξ ðy 0 Þ = 1/2, and F ξ ðγ + y 0 Þ = 3/4, the former formula states that y 0 is the median of Cðγ, y 0 Þ, while the latter formula states that γ + y 0 is the 3/4 quantile of Cðγ, y 0 Þ.
In particular, the Cauchy distribution with parameter y 0 = 0, γ = 1 is called the standard Cauchy distribution, denoted by ξ~Cðγ, y 0 Þ. Its probability density function is Its cumulative distribution function is The properties of the Cauchy distribution: the expectation and variance are undefined, the median or mode is y 0 , the entropy value is ln 4πx, and the eigenfunction is exp ðy 0 it − γjtjÞ.
It can be seen from the characteristics of the Cauchy distribution that the Cauchy distribution is a continuous distribution function with neither expectation nor variance. Therefore, the classical parameter estimation method cannot be used to estimate the parameters of the Cauchy distribution [17].
For continuous distribution functions, the inverse cumulative distribution function can be used for sampling. 4 Wireless Communications and Mobile Computing Supposing the distribution function of the sample random variable X is FðXÞ, and the function FðXÞ is continuous. Its inverse cumulative distribution function is also its inverse function, denoted as F −1 ðzÞ, 0 ≤ z ≤ 1.
In Theorem 1, if the random variable R is uniformly distributed on (0,1), that is, R~Uð0, 1Þ, then Y = F −1 ðRÞ is the inverse function of FðYÞ.
From the cumulative distribution function formula of the Cauchy distribution, its inverse cumulative distribution function can be obtained as It can be known from Theorem 1 that if people want to generate a random variable number that obeys FðYÞ, first generate a random variable number u that obeys U(0,1) uniform distribution, and finally calculate F −1 ðuÞ.
The test function SumCan is shaped like an inverted spire, and the optimal position is at this inverted spire, and other positions are relatively flat. Therefore, the position of the global optimal solution searched by many algorithms during optimization is very close to the optimal position of the test function, but it is very different from the optimal fitness value, which is -10 5 .
The test function Sphere is a unimodal function with simple relationships between variables. The function test Schwefel is a multimodal function, and its local optimum position is far away from its global optimum position because its local optimum value is also very small. Thus, when it converges to the local optimum, it is easy to mistakenly think that it has converged to the global optimum. At the same time, it is difficult for the algorithm to get rid of the local optimum that has been trapped, and it is difficult to achieve the level of global optimization. Therefore, this test function is often used to analyze the execution performance of algorithms out of local optima. The test function Rastrigin is a multimodal function. Similar to the test function Schwefel, it is used to test the execution ability of the algorithm to get rid of the local optimum. The difference is that its local optimum is located near the global optimum. So it is easy to fall into local optimum. The test function Griewank is a multimodal function with many local optimal points, and each dimension variable has a strong interaction relationship. Due to the strong interaction and influence of each dimension variable, the number of local optimal points will increase with the increase of dimensions and increase sharply, thus making it extremely easy for the algorithm to fall into local optimum points and not easily get out of it. Therefore, the test function is often used to test the ability of the algorithm to get rid of local optima and the performance of the optimization problem for high-dimensional variables. The optimal solution and optimal value of the test function are shown in Table 1.
The simulation results of the test function in the marginal distribution are shown in Figure 4.
The population size is 2000, the dimension is 10, and the selection rate is 0.5. Some individuals are selected as part of the new population by truncation selection and roulette selection. The mutation operator is used in the algorithm, and the mutation rate is 0.05. The algorithm stops when any of the following conditions are met. The conditions are as follows: ① The difference between the optimal values of the evolution of the adjacent two generations of the algorithm within 25 consecutive generations is less than 1e-6; ② the optimal value is found; and ③ the maximum fitness evaluation times are 300,000. Each function was run 50 times independently in different test environments. In From the experimental results in Table 2, it can be seen that for the four test functions, the copulaEDA whose mar-ginal distribution is a Cauchy distribution function is better than the copulaEDA with an empirical distribution function. For the test function Rosenbrock, it can be seen from Table 2 that whether the empirical distribution function is used as the marginal distribution or the Cauchy distribution function is used as the marginal distribution function, the Relatively speaking, the copula distribution estimation algorithm of the Cauchy distribution probability model is relatively better than the copula distribution estimation algorithm of the empirical distribution probability model in terms of optimization performance and tending to stability [19]. For the test function SumCan, it is also a difficult problem in the distribution estimation algorithm to continue to optimize. From the results in Table 2, it can be seen that the optimization results of the two algorithms are not good and the difference is not large. For the test function Sphere, it can be seen from the experimental results that the optimization result of the copula distribution estimation algorithm of the Cauchy distribution probability model on this function is better than the copula distribution estimation algorithm of the empirical distribution probability model. For the test function Schwefel, the copula distribution estimation algorithm of the Cauchy probability model is obviously superior to the Clayton copula distribution estimation algorithm of the empirical distribution probability model in terms of mean, variance, minimum and maximum values during optimization. After careful calculation, the optimization rate of the Cauchy model has reached more than 10%.
Gesture Recognition Robustness Verification.
In the current vision-based gesture recognition, most of them use color cameras to obtain image information, while the color camera-based visual gesture recognition system has high requirements on the environment. And it is easily affected by changes in illumination, resulting in the inability to accurately segment gestures part, which ultimately affects the result of gesture recognition; in skin-color-based gesture detection, it is unavoidable that it will be affected by the skin-like environment and its own skin color and gestures cannot be extracted from complex backgrounds. This paper uses the Kinect depth camera to obtain RGB-D image information. The infrared camera group it contains can project and receive infrared light to obtain depth data within the visible environment. Changes in ambient lighting will not affect the acquisition of depth data, and depth information can be used for gesture recognition even in dark environments [20]. Through the depth threshold gesture segmentation fused with skin color information, the interference of non-skin color environmental factors in the effective gesture area can be excluded, thereby improving the accuracy of gesture segmentation. In the calculation of the gesture feature parameters, the judgment of the effective convex defect and the ratio of the gesture area to the minimum circumscribed circle area are all in the form of ratio, which makes the gesture recognition have good adaptability; the effective convex defect feature angle and the effective vertex itself have scaling and rotation invariance, correct recognition of gesture rotation, and scaling. In order to verify the robustness of the static gesture recognition method, gesture recognition experiments are carried out under strong light, weak light, and Table 3.
The static gesture recognition method adopted in this paper can well overcome the influence of different light intensities and can also correctly recognize gestures in the presence of background interference [21]. The experimental results of gesture recognition under different operators and gesture rotation and zoom conditions are shown in Table 4. Although different operators have differences in expressing gestures, or excessive zooming in the process of gesture changes weakens gesture features, it will eventually lead to identify the wrong situation, but can get a good recognition effect within the effective recognition range.
An in-depth study on static gesture recognition based on multiparameter features from image acquisition to gesture recognition have been done [22]. The RGB-D image information is obtained by using the Kinect depth sensor, and then the gesture image is segmented by the adaptive depth threshold fused with skin color, and the corresponding preprocessing is performed to extract the outline of the gesture. Experiments show that the static gesture recognition method used in this paper is simple and high in recognition and has good robustness, and the accuracy rate can basically reach more than 90%. Figure 5 clearly describes the entire experimental process.
Experimental Process. The experimental flow chart in
First, selecting the optimal threshold value, fully consider the accuracy of the recommendation achieved by the threshold value and the time-consuming situation, and considering the effect of the threshold value in the subsequent selection of the number of neighbors. Selecting the optimal threshold to achieve gray project clustering makes the internal common features of the project prominent, and the external differences between the projects are large [23]. Calculating the similarity of items on the obtained clusters, and according to the fixed number of neighbors, sort the calculated similarities from large to small in turn, and select a fixed number of neighbors. The scoring prediction and recommendation of the item are completed, and the performance of the recommendation algorithm is verified experimentally.
Performance Evaluation
Criteria. The performance of the algorithm is verified by experimental simulation. During the experiment, other recommendation algorithms are used to compare and analyze the algorithm. The movie data set is used, and the recommendation algorithm is used to operate the data set. And the obtained simulation results can test the efficiency and accuracy of the algorithm we study.
The evaluation method of the pros and cons of the recommendation algorithm has also been under exploration. Nowadays, the commonly used evaluation schemes are as follows: decision support accuracy and statistical accuracy. The most commonly used is the mean absolute difference measurement method among the statistical accuracy measurement methods. Therefore, this paper uses the MAE algorithm to evaluate the recommendation system. The MAE value represents the value of the degree of deviation between the predicted user score calculated by the algorithm and the actual score, so when the obtained value is smaller, it means that the difference between the predicted value and the actual value is smaller, so it also proves that the accuracy of the recommendation algorithm is higher, which also clarifies the effectiveness of the recommendation algorithm [24].
The calculation formula of the MAE evaluation scheme is as follows: In Formula (23), G j represents the predicted value of the user's rating, A j represents the actual value of the known user's rating, and T represents the total number of rating items.
Selection of Optimal Threshold for Grey Clustering.
Given a known movie data set, the classification situation has been clarified, and the original data set is subjected to Items with similarity within the threshold range clustering to get N clusters User similarity calculation within the same cluster. Get similarity matrix S Using the K nearest neighbor method to select the optimal nearest neighbor, to the nearest neighbor set of corresponding users Make the score and prediction of the project based on the user nearest neighbor set Performance evaluation of recommended projects Wireless Communications and Mobile Computing the gray theory to realize the project clustering process, and a new data set classification situation will be obtained, and the gray project clustering algorithm is based on thresholds are used to classify categories. Therefore, in the experimental process, the known classification situation is compared with the clustering results obtained by the gray clustering algorithm when the threshold values are different so that it can be clear that the highest accuracy obtained under a certain clustering threshold can be obtained the classification result. During the experiment, the threshold value is from 0.1 to 0.9, and the obtained clustering result is shown in Figure 6, which vividly describes the change trend of the clustering accuracy curve under the condition of different threshold selection [25]. It can be clearly seen from Figure 6 that in the process of obtaining the threshold from 0.1 to 0.7, the accuracy of the item clustering division of the gray correlation algorithm is a monotonically increasing function, and the accuracy is known to improve; when the threshold value is obtained from 0.7 to 0.9, it can be clearly seen from the vertical coordinate of the accuracy in Figure 6 that the accuracy has a downward trend. It can be seen that when the grey relational clustering threshold is 0.7, the highest clustering accuracy can be obtained, that is, the divided item clusters are the least different from the actual known item clusters [26]. Therefore, the optimal threshold of the grey relational clustering algorithm obtained is 0.7. Table 5 describes the number of samples of clustering errors generated under each threshold and the corresponding clustering accuracy when different clustering thresholds are selected. And the number of categories generated by clustering has made a specific numerical elaboration and comparison.
It can be seen from Table 5 that when the optimal threshold for grey relational clustering is 0.7, the accuracy of the clustering algorithm reaches 73.3%, and a total of 121 categories of movie clusters are generated. In order to further optimize the selection of the threshold, the experimental range of the threshold is selected between 0.6 and 0.8, and the obtained clustering accuracy is shown in Figure 7.
From Figure 7, we can see that the accuracy of clustering shows a clear upward trend in the process of changing the threshold from 0.60 to 0.75, while the accuracy decreases Tables 6 and 7 describe the clustering accuracy obtained with further refinement of the clustering threshold selection. Table 6 describes the clustering accuracy analysis when the threshold range is between 0.61 and 0.69, and Table 7 describes the accuracy when the clustering threshold is selected between 0.71 and 0.79. The number of samples corresponding to clustering errors under each threshold and the corresponding clustering accuracy can be used to determine the optimal threshold more accurately [27]. Table 7 is an analysis of the clustering results of items with a threshold in the range of 0.71-0.79.
It can be seen from Tables 6 and 7 that when the gray correlation clustering threshold is 0.75, the highest clustering accuracy can be obtained, that is, the difference between the divided item clusters and the actual known item clusters is the least, so the optimal threshold of the obtained grey relational clustering algorithm is 0.75.
In the grey relational clustering algorithm in this paper, the clustering based on user items is completed, so in the process of comparing recommendation algorithms, it cannot be compared with the commonly used classical clustering methods that use item attributes as a reference. Therefore, in order to better study the performance of the recommendation algorithm, we introduce a gray-based collaborative filtering recommendation algorithm for comparison to verify the effectiveness of gray clustering. The ratio of the test set to the training set is 1 : 4 randomly divided, 20% of the test set and 80% of the training set are selected, and the number of neighbors is 40. Under different gray clustering thresholds, the item scores are predicted and the MAE value is obtained. It can be seen from Figure 8 that under different threshold values, when the gray correlation threshold is 0.75, the MAE value reaches the minimum value, which means that the gap between the predicted score and the actual score is the smallest, which means that the predicted result is very accurate, so it can be proved that the recommendation algorithm has relatively superior performance. Therefore, choosing the optimal threshold value as 0.75 can achieve the best user recommendation effect.
Discussion
Although the method in this paper reduces the complexity of the system to a certain extent, if the system is applied in an embedded platform, the running speed is still relatively slow. How to further improve the recognition efficiency of the whole system is the focus of future research. In the process of threshold selection for grey item clustering, several experiments are carried out in this paper to obtain a fixed optimal threshold. In the future, the dynamic adaptive algorithm of item clustering threshold can be studied. In the process of finding the optimal threshold, the optimal threshold is used to achieve the highest accuracy of gray item clustering and to alleviate the influence of the number of neighbors on the selection of neighbors and the final recommendation effect.
Conclusions
In edge computing, the use of Cauchy distribution processing is more impressive; compared with empirical distribution, the algorithm optimization can reach at least 10%.
Experiments show that the static gesture recognition method used in this paper is simple and high in recognition and has good robustness and the accuracy rate can basically reach more than 90%. In the case of different threshold values, when the gray correlation threshold is 0.75, the MAE value reaches the minimum value, which means that the gap between the predicted score and the actual score is the smallest, which means that the predicted result is accurate, which can prove that the recommendation of the algorithm has relatively superior performance. Therefore, the optimal threshold value is selected as 0.75, which can achieve the best effect. Moreover, aiming at the behavior of the manipulator, we will have a deeper understanding of computer-related technologies and simulate a better behavior of the manipulator.
Data Availability
Data sharing is not applicable to this article as no new data was created or analyzed in this study.
Conflicts of Interest
The author states that this article has no conflict of interest. | 7,449.4 | 2022-08-09T00:00:00.000 | [
"Computer Science"
] |
A Systematic Review of Construction 4.0 in the Context of the BIM 4.0 Premise
: This paper presents a systematic review of Construction 4.0 in the context of the building information modeling (BIM) 4.0 premise. It comprises a review of the industry in the pre-fourth industrial revolution (4IR) age, the current and anticipated development of the 4IR, Construction 4.0’s origin and applications, and the synergy of its main drivers, i.e., the synergy of BIM with the internet of things (IoT) and big data (BD). The main aim of the paper is to determine the Construction 4.0 drivers and to what extent are they initialized by the 4IR, their development and their synergy with BIM, and the direction of BIM’s implementation in the construction phase. It was found that the main drivers of Construction 4.0, which originated from the 4IR, are BIM, IoT, and BD, but with specific implementations. The results of the analysis of BIM with IoT and/or BD revealed that the integrative approaches combining the aforementioned drivers show signs of project enhancement by providing significant benefits, such as improved real-time monitoring, data exchange and analysis, construction planning, and modeling. Furthermore, it was revealed that the main drivers are mostly applied in the project’s preconstruction phase, which is continuously developing and becoming more automated. The state-of-the-art review presented in this paper suggests that BIM is in transition, adopting Construction 4.0 to become BIM 4.0.
Introduction
The fourth industrial revolution (4IR), also colloquially referred to as Industry 4.0, is expected to bring growth, enhancement, and accelerated development to most industries in the near future. In comparison to the previous technological revolutions, the 4IR could be the first revolution simultaneously active in most parts of the world, due to globalization trends. Industries that have already stepped in and adopted the 4IR report that 4IR changes are mostly stimulated by the emergence of new technologies, i.e., drivers, such as digital twin construction (DTC), building information modeling (BIM), internet of things (IoT), Big Data (BD), and additive manufacturing (AM)/3D printing. These technologies have, to a certain extent, already changed most industries, but industries are still not seriously adopting the full potential of the 4IR. An example of such partial implementation of the aforementioned technologies is the construction industry, which in light of the 4IR is often referred to as Construction 4.0. The introduction of the Construction 4.0 concept and new technologies is anticipated to be a major challenge for a commonly sluggish industry. Among numerous recent reports and strategic studies regarding Construction 4.0, a report published in 2016 by the Roland Berger consultant company [1] stated that Construction 4.0 provides a variety of possibilities for stakeholders in the construction industry to boost their productivity in all kinds of ways. However, just 6% of construction companies of construction companies make full use of digital planning tools, while 93% of them agree that digitization will affect every process. Despite the poor adoption rates, a report from 2019 [2], published by the Publications Office of the European Union, underlines the potential of digital transformation in the architecture, engineering, and construction (AEC) sector as
Research Methods
This paper presents a systematic literature review, with references filtered and extracted from recently published relevant scientific papers, reports, and conference papers indexed in Web of Science and Scopus. Since the topic is developing worldwide, there have been numerous studies published in the past decade. The research was initialized by determining the main keywords (i.e., third industrial revolution, fourth industrial revolution, industry 4.0, construction 4.0, cyberphysical systems, CPS, digital twins, BIM, BIM 4.0, internet of things, IoT, Big Data, additive manufacturing, and 3D printing), which were used in various combinations for each segment of the paper. In Figure 1, the scope and gradual importance of research objectives are presented. Since the topics, i.e., the foci of the research, overlap and can hardly be thematically or periodically discretely separated, their overlaps are presented as fuzzy, as shown in Figure 1.
Due to the fact that the topics of the research move from a wide scope to a single research objective, the main challenge was filtering the relevant literature. In our first literature search, the keywords highlighted in the systematic literature review by Oesterreich and Teuteberg [4] were applied, and we went through several rounds of narrowing, aiming at simultaneous appearances of both BIM and BD, as well as BIM and IoT. The vast majority of references date from 2010-2020, with a few exceptions in the first chapter regarding previous industrial revolutions. The literature review presented in this research resulted in a total of 172 referenced sources. A standard software tool for constructing and visualizing bibliometric networks, i.e., VOS viewer software, was used to present the keywords referenced in the paper [23]. Due to the fact that the topics of the research move from a wide scope to a single research objective, the main challenge was filtering the relevant literature. In our first literature search, the keywords highlighted in the systematic literature review by Oesterreich and Teuteberg [4] were applied, and we went through several rounds of narrowing, aiming at simultaneous appearances of both BIM and BD, as well as BIM and IoT. The vast majority of references date from 2010-2020, with a few exceptions in the first chapter regarding previous industrial revolutions.
The literature review presented in this research resulted in a total of 172 referenced sources. A standard software tool for constructing and visualizing bibliometric networks, i.e., VOS viewer software, was used to present the keywords referenced in the paper [23].
Industries Pre-4IR
History has shown that industrial revolutions tend to have a slow starting pace, but with time, have a galloping impact on shaping common production technologies, everyday lifestyles, etc. [24]. In general, the term industrial revolution can be defined as a widespread dramatic change in the methods of producing goods and services [25]. Like the previous, i.e., first and second industrial revolutions, the third industrial revolution (3IR) was also driven by technological advances regarding manufacturing, distribution, and energy factors [24]. In the first industrial revolution, it was the printing press, in the second industrial revolution it was radio and television, and in the 3IR, it is/was the combined power of computing, telecommunications, and news broadcasting [26]. It is believed that society will emerge from the 3IR as a dynamic "global village" because technology companies, content providers, and information professionals will empower people to browse, retrieve, share, and use data for personal and professional uses. While the 3IR indeed fulfilled most of its potentials in the majority of the developed countries, in developing economies it still has not [25]. Digital tools and equipment are still becoming widely used for either designing or manufacturing products enhancing the sharing of designs and easier collaboration among stakeholders. Therefore, the manufacturing resources pool is significantly larger in scale than what any single maker could achieve [27]. On the other hand, it is believed that direct digital manufacturing is not merely a stimulus of the 3IR, but one of its effects. The main challenge of the 3IR was found to be the traditionalism of most industries, manifested in the sluggish upgrade of established enterprises in accepting and implementing reengineering [28]. When it comes to 3IR technologies, among others, six major high-technology agents are underlined in the literature, i.e., mi-
Industries Pre-4IR
History has shown that industrial revolutions tend to have a slow starting pace, but with time, have a galloping impact on shaping common production technologies, everyday lifestyles, etc. [24]. In general, the term industrial revolution can be defined as a widespread dramatic change in the methods of producing goods and services [25]. Like the previous, i.e., first and second industrial revolutions, the third industrial revolution (3IR) was also driven by technological advances regarding manufacturing, distribution, and energy factors [24]. In the first industrial revolution, it was the printing press, in the second industrial revolution it was radio and television, and in the 3IR, it is/was the combined power of computing, telecommunications, and news broadcasting [26]. It is believed that society will emerge from the 3IR as a dynamic "global village" because technology companies, content providers, and information professionals will empower people to browse, retrieve, share, and use data for personal and professional uses. While the 3IR indeed fulfilled most of its potentials in the majority of the developed countries, in developing economies it still has not [25]. Digital tools and equipment are still becoming widely used for either designing or manufacturing products enhancing the sharing of designs and easier collaboration among stakeholders. Therefore, the manufacturing resources pool is significantly larger in scale than what any single maker could achieve [27]. On the other hand, it is believed that direct digital manufacturing is not merely a stimulus of the 3IR, but one of its effects. The main challenge of the 3IR was found to be the traditionalism of most industries, manifested in the sluggish upgrade of established enterprises in accepting and implementing reengineering [28]. When it comes to 3IR technologies, among others, six major high-technology agents are underlined in the literature, i.e., microprocessor, computer-aided design and manufacturing (CAD/CAM), fiber optics, biogenetics, lasers, and holography [28]. There is a special emphasis on the development of microelectronics technology at this historical juncture [29]. The main reason for its importance is an immense impact on the affordability of computing power. Due to the simultaneous reduction in the cost of computers and the massive increase in their power, the microprocessor made computers accessible to a large number of people who could not have afforded or operated their predecessors [30]. An additional challenge is the interaction between technological changes and the international division of labor [29]. The 3IR began to affect labor in industrialized countries by the late 1970s because, in developed countries, increased income and living standards made customers more sophisticated and demanding. At the same time, the market for lasting consumer goods was saturated and their demand subsided. As for the negative side effects, unemployment arose during the 3IR and is mostly associated with difficulties in the process of transferring the labor force from industry to services [31]. Additionally, the economic migration of many workers to more developed countries has caused shortages in the workforce in less developed countries, which will continue to persist as a problem in the 4IR [32]. It is anticipated that some of the problems and challenges that the 3IR faced will be resolved in the 4IR, but new ones may emerge.
The Fourth Industrial Revolution (4IR)
Even though some benefits of the 3IR have not yet reached much of the world's population, in developed countries the 4IR has already taken its place [33]. The 4IR, also referred to as "Industry 4.0," made its first appearance at the Hannover Fair in 2011 [34]. Schwab [33] characterized the 4IR as "a fusion of technologies that are blurring the lines between the physical, digital, and biological spheres." Unlike the previous Industrial Revolutions, the 4IR is progressing at an exponential pace, not sluggishly nor in a linear manner. It is expected that, in the future, technological innovation will reinforce the supply side and bring about gains in efficiency and productivity in the long run. Additionally, production process automation aims to reduce the scale problem of labor force deficiency reported in the 3IR. It is believed that it is necessary to implement new technologies for automation to achieve complete digitization and intelligence of existing industrial processes [35]. Therefore, the future of manufacturing may see industrial production systems become more intelligent by using digital systems to create more knowledge-based productions, which will greatly improve their efficiency and competitiveness.
As described above, the 4IR is considered to be mainly dependent on building a CPS to create a digital and intelligent factory, to navigate manufacturing towards becoming more digital, information-led, customized, and sustainable [36]. The 4IR integrates IT systems with physical systems to get a CPS that brings the real world into VR [37]. Those systems represent the integration of an information system (IT) with mechanical and electronic components that are connected to online networks and allow for communication between machines in a way that is similar to social networks [38]. The cyberphysical integration is also enabled by the digital twin (DT) concept, which can be considered a necessary path to realize CPS [39]. Ultimately, CPSs and DTs enable the integration of production, sustainability, and customer satisfaction while forming the basis of intelligent network systems and processes [40]. Besides CPS and DT, 4IR also uses IoT to connect production technologies with smart production processes to make manufacturing smart [41,42]. The basic idea of IoT is to make "things" around us communicate with each other to achieve mutual goals, with its main feature being the integration of various identification and tracking technologies, i.e., wired and wireless sensors and actuator networks, enhanced communication protocols, and distributed intelligence for smart objects [43]. The implementation of the IoT concept will be enhanced by the fifth-generation mobile network (5G), which is the term used to describe the next generation of wireless networks. The features of the 5G network will provide the user with several performance enhancements regarding network capacity increase, shorter latency, more mobility, and increased network reliability and security, which will, in turn, result in an all-connected environment called the IoT [44]. The final puzzle is the structure or environment that can handle the managed information by CPS and IoT, and that is BD and cloud computing [45]. With that being said, it is clear that one of the most important technologies, besides IoT, adopted in the 4IR is BD, which is related to the collection, processing, and analysis of a large amount of structured and unstructured data with intelligent algorithms [37]. The term BD is derived from the fact that the datasets are so large that typical database systems are not able to store and analyze them; also, the data are no longer traditionally structured, but originate from many new sources including e-mail, social media, and Internet-accessible sensors [46]. Using BD to replace processes that are done manually may make certain jobs outdated, but may also create new categories of jobs and opportunities that currently do not exist in the market [33]. One of the definitions of BD provided in [47], and also the most widespread one, is the "4V" theory, stating that BD comprises a variety of resources and contains a great volume of data; BD streams in at a high rate and must be handled timely, which implies velocity; BD comes in a variety of formats; BD has to be cleaned to ensure the validity. BD analysis may require a considerable commitment of hardware using the old hardware storage method, but the emergence of cloud computing promises to make it small, by reducing computational costs while increasing the elasticity and reliability of systems [48]. Another important feature of the 4IR is AM or 3D printing, which represents the capability of producing three-dimensional objects from virtual models [45]. According to [49], the advantages of AM are the possibility of designing and developing products. Additionally, companies are using AM to capitalize on its benefits like complexity for free manufacturing, while in traditional manufacturing a direct connection between complexity and manufacturing costs exists. The aforementioned technologies are becoming increasingly implemented in many industries, including in the construction industry, where the whole concept has merited a new term, Construction 4.0.
Construction 4.0 4.1. The Construction 4.0 Paradigm
It is a common belief that the first mention of Construction 4.0 dates back to 2016, and was primarily based on construction companies' awareness of the importance of digitization in the construction industry [1,50]. Thus, it can be said that Construction 4.0 is the convergence of industrial production, CPSs, and digital technologies with the ultimate goal of creating a digital construction site [34]. As such, it is anticipated that Construction 4.0 will fundamentally influence organizational and project structures, since the framework of Construction 4.0 enables planning, designing, and delivering built assets more effectively and efficiently, with the focus being on the physical-to-digital transformation and then digital-to-physical [51,52]. Construction 4.0 can be defined as a paradigm that comprises CPSs and the internet of things, data, and services, with the main aim of connecting the digital layer, which consists of BIM and the common data environment (CDE), with the physical layer, which consists of the asset and its lifecycle. Besides CPS, the Construction 4.0 framework also uses digital ecosystems and links them with CPS, which is used as a core driver [53], where digital ecosystems represent "an interdependent group of enterprises, people, and/or things that share standardized digital platforms for a mutually beneficial purpose, such as commercial gain, innovation or common interest." The conceptual model of a digital ecosystem consists of a business network of third-party developers, boundary objects, and a core digital platform [34]. The ultimate goal is to create an interconnected environment that integrates organizations, processes, and information with the purpose of efficiently designing, constructing, and operating assets [54]. According to a report of the Digital Supply Chains in the Built Environment Work Group (DSCiBE) [55], the introduction of BIM can be considered the first step towards a collaborative digital communication and has also pushed the construction industry to look at how it can deliver value through data. The main aim of the report by the DSCiBE task group was the standardization and interoperability of product data as well as digital product identification. As expected, the drivers of Construction 4.0 have their benefits and challenges; the main ones are presented in Table 1. adoption of the lifecycle building approach, reduction of waste and efficiency improvement, horizontal, vertical, and longitudinal integration, improving sustainability, cost and time reduction, improved safety performance, enhanced quality of buildings, improvement of the poor image of the construction industry high initial investments, lack of skilled workforce and the need for enhanced work skills, deficiency of globally agreed standards for the construction industry, data security, i.e., cybersecurity lack of knowledge about Construction 4.0 resistance of the construction industry to change Identification of the Construction 4.0 drivers, as the initial step in its development, has been a relevant research topic in the last five years. Various authors have identified various drivers, i.e., various technologies that have enabled the emergence of the Construction 4.0 concept. In 2016, popular media such as newspaper articles, magazine articles, blogs, and websites were analyzed to determine which technologies are considered a part of the 4IR. It was found that the central technologies are BIM, Cloud Computing, and IoT. Moreover, it was concluded that all of the 4IR technologies are at different levels of maturity [4]. In 2019, research presented in [63] determined that there is an active collaboration between BIM and 4IR technologies. Additionally, it was found that there is a lack of understanding of the 4IR concept in the construction industry [63]. In 2020, four technologies were determined to be essential to the understanding of Construction 4.0: 3D printing, BD, VR, and IoT. The research was conducted using a bibliometric analysis and by analyzing the keyword occurrence [50]. In the same year, Maskuriy et al. [64] researched the application of 4IR technologies in construction and found that most integrated technologies have focused on the preconstruction phase. Furthermore, in the literature review in [65], it was noted that Germany leads the field of Construction 4.0, and is followed by China, the United States, etc. Moreover, a UK-based multinational construction design consultancy firm was analyzed, and the results showed that there are many barriers to implementing Construction 4.0, such as residual managerial practices. As some of the main enablers for the implementation of the Construction 4.0 concept, the following technologies were identified: IoT, Cloud computing, BD, Artificial intelligence (AI) and robotics, and cybersecurity [65]. Table 2 presents the summarized chronological findings and applied methodologies regarding Construction 4.0 drivers in the last five years. It can be noted that the aforementioned challenges of Construction 4.0 have also been reported by the authors presented in Table 2. Additionally, the most prevalent drivers of Construction 4.0 are considered to be BIM, IoT, and BD, which are mentioned by most of the authors. The stimulus for applying CPS in the construction industry is their integration of physical systems and their virtual representations, i.e., their DTs, to create an integrated analytical system, where a DT is the real-time digital representation of a building or infrastructure [66]. Such a system should be able to adapt to changes at construction sites and connect the virtual world with the physical world by using sensors or data acquisition technologies and actuators [67,68]. The changes should update the state in the form of measurements, data, and pictures, which are then updated in the DT and allow for continuous monitoring of a 4D BIM model [66]. According to [34], CPS consists of the physical part, which is usually a device, a machine, or building, and a cyber part, which is usually data, a software system, or a communication network. Furthermore, CPSs are systems of interconnected physical and digital twins, where the digital twins are the virtual assets or simulations of the physical object in real time. Additionally, digital and physical twins are reciprocally connected by sensors and actuators. By enabling a tight connection between computational models and associated physical entities, the integration of CPS and DT offers a way for construction project teams to bridge the gap between virtual models and physical construction and can therefore be considered the "heart" of Construction 4.0 [58]. Many innovative technologies, such as prefabrication, automation, 3D printing, VR, AR, unmanned aerial vehicles (UAV), sensor networks, and robotics for repetitive or unsafe procedures, are enabled by bidirectional communication between construction components and their digital representations [34,58]. Therefore, the built environment is a rich area for the application of the CPS and DT framework since smart buildings, cities, and infrastructure are all examples of what may be called cyberphysical environments, where the built environment becomes increasingly intelligent and digitally connected [69]. Ultimately, CPS can be considered as the key to achieving more efficient, safer, and more environmentally friendly construction projects, which are also the goals of Construction 4.0 [34]. As mentioned earlier, CPS consists of two principal elements, i.e., the "physical to cyber" bridge and the "cyber to physical" bridge [70,71]. In terms of construction, the physical to cyber bridge represents construction components and processes that are tracked using sensors and other tracking systems [58]. In addition, the progress and changes in the construction process are monitored and coordinated with their associated cyber representations for further action. The cyber to physical bridge covers the actuation dimension and dictates how the information from the sensors is used to manage the system, which means that actuators in this sense involve transmitting appropriate information to enable prompt decision making. Improved safety has the potential to be the key benefit of implementing CPS and DT in the construction industry since it is predicted that project managers and safety specialists will have access to locations of employees and heavy equipment at all times [72,73]. Additionally, sensors have made data exchange among workers easier and provided opportunities to monitor their health to increase safety [74]. Moreover, structure monitoring sensors can detect malfunctioning structural components to ensure site safety.
In [58], the application of CPS for various purposes in the construction industry was analyzed, such as for construction component tracking, temporary structure monitoring, and mobile crane safety. Furthermore, a cyber model named Petri Net [75] was adopted as a model for a construction process and two application scenarios of automatic assembly and traditional structural masonry were simulated. In [76], a citizen service center was presented for the verification of the technical feasibility and implementation effect of a CPS framework, in which the real-time construction model acts as the digital twin of the building under construction. It was concluded that, benefitting from real-time monitoring, simulations, and the decision support mechanism of the proposed CPS, future construction plans will no longer be fixed and predefined, but it will be possible to make and adjust schedules according to the actual situation in the construction process.
Building Information Modeling (BIM)
The subject of BIM currently represents a central topic for the improvement of the construction industry, and simultaneously a core technology for supporting the idea of the 4IR in the construction industry [4]. Since BIM is increasingly being adopted and utilized in the architecture, engineering, and construction (AEC) domain [77], it can act as a catalyst for deeper adoption of digitization, which is also due to the quick and good acceptance it has received [78]. As the center of the digitization of the construction industry, together with the 4IR concept, BIM can close the digital gap that still exists and have a positive impact on future building processes [79]. By incorporating various properties, BIM can offer a high-accuracy representation of a project at the level of components [76], and an integrated three-dimensional model can be adopted to completely express the definition information of buildings [80]. An important feature is the bidirectional coordination between the physical and virtual domains. That coordination leads to a digital replica of the building, which improves the control and optimization of the construction process while also generating valuable data for the building's operation/maintenance, as well as for the design and planning phase of future construction [63]. The mentioned virtual replica of the building can be compared to a similar concept in the manufacturing industry called the digital twin [81]. Various subsets of BIM can be referred to as dimensions, where 3D is the object model, 4D is time, 5D is cost, 6D is operation, 7D is sustainability, and 8D is safety [59].
Among others, an interesting study that represents a notable example of the importance of BIM by pointing out that BIM can help users share project information during the entire construction lifecycle was presented in [82]. The building's information is available to everyone included in the construction project, from the design team to the construction team and the owner of the building. Additionally, all mentioned project members have the possibility of adding or changing information during their period of using the BIM model, which is done by the integration of BIM into cloud computing. Thus, the project stakeholders can collaborate in real time from different locations to enhance decision-making and ensure project deliverability [83]. BIM is also useful in terms of reducing the data size, because the volume of data collected from construction projects is massive due to the complexity of their designs and construction activities [84]. However, with the support of BIM, the volume of data that is related to the design of a three-story building can easily reach 50 GB [85], which burdens the data-interoperability and data-transfer processes.
Alongside all the benefits stated, it is important to mention the main challenge that BIM faces, which is related to the full implementation of this technology in all phases of the design, construction, and operation of buildings [86]. Moreover, the biggest constraints on the mainstream application of BIM refer to the lack of scalability, interoperability, and support for remote collaboration [87,88]. Several studies have been carried out on the use of BIM in the construction industry, e.g., on cloud-BIM [83] and also linking BIM to construction lifecycle phases [89]. Furthermore, a paper presented the linkage between sensor and BIM by using IFC (industry foundation classes) [90]. Additionally, research was conducted on the development of construction industrialization based on BIM methodology, and a navigation framework was proposed for the inclusion of BIM and factory equipment to simulate a digital twin factory [91]. Another interesting paper is [92], wherein the potential of BIM and the 4IR to change the future of the construction environment was discussed.
In the context of recent trends of BIM development, it is important to mention integrated building information modeling (iBIM). According to [93], the main aim of iBIM is to integrate BIM with other innovative technologies and managerial approaches during the project life cycle. To enhance the project performance, BIM can be integrated into three stages of the project: the preconstruction phase, the construction phase, and the facility management phase. The preconstruction phase concerns the integration of BIM with materials tracking and logistics systems to benefit from supply chain management [93]. An example of such integration is presented in [94], in which the potential of BIM to provide contextual information mapping across processes was observed. It was also stated that the integration of design, manufacturing, and construction processes, with transparency of information about material resources across these processes, would bring about significant benefits for all stakeholders within the supply chain. The integration of BIM and GIS was used in the planning and designing phase of the building for activities such as construction site selection, energy design, structural design, performance evaluation, etc. [95]. Another interesting example of integrating BIM and GIS in the planning phase of construction is a case study that presented a BIM-GIS system for visualizing the supply chain process and the actual status of materials through the supply chain [96]. Furthermore, multiple studies regarding the integration of BIM and RFID for various purposes were conducted [97][98][99][100]. In the construction phase, conceptual frameworks have been researched for the integration of BIM and AR [101,102]. It was investigated how BIM can be extended to the site via AR to improve the way information is accessed [103]. Furthermore, BIM can also contribute to quality management by integration with Light Detection and Ranging (LiDAR), with the purpose of better evaluation of on-site conditions and achieving real-time construction quality control [104]. The last phase is the maintenance/operation phase, in which the integration of BIM and LiDAR has also found its application; an example is the automated inspection of the quality of prefabricated houses [105]. This phase reflects the integration of BIM and AR, which is extremely useful in the field of facility management; an example is a BIM2MAR method tested in a facility management pilot study [106]. The versatile possibilities of BIM application in the construction industry could be the foundation for the change that the construction industry needs.
Internet of Things (IoT)
A technology that brings physical objects into a cyber world that is based on devices or technology such as sensors, actuators, RFID, video cameras, and laser scanners is called the IoT [76,107]. According to [108], the IoT is a set of four different layers, i.e., an application layer, a perception layer, a network layer, and a physical layer. The layer that refers to smart cities, smart transport, and intelligent homes is the application layer. The perception layer refers to technologies that communicate with other objects, like sensors and devices. Furthermore, the network layer refers to the network communication and the component of network coverage, while the physical layer refers to the hardware and includes smart appliances and other devices. The use of the IoT in the construction industry has many possibilities and benefits, mostly focused on fast decision making due to the availability of real-time data analytics [109,110]. In addition, IoT technologies and applications could transform the construction, maintenance, and operation phases by maximizing user comfort, security, and energy saving by diverse intelligent solutions [111]. However, it is believed that most current IoT solutions in the construction industry are isolated for specific applications but lack coordination over the entire construction process [76]. Research carried out and presented in [112] determined the dominant challenges to applying IoT in the construction industry. It was found that these challenges are a lack of safety and security, a lack of documented standards, a lack of awareness of the benefits, the improper introduction of IoT, and a lack of robustness in connectivity. These challenges have also been reported in other papers [113][114][115].
Despite the mentioned challenges, the IoT has been widely accepted in the construction industry. For example, by taking advantage of the IoT, the real-time data collected from a construction site drives BIM models to monitor the construction process [116]. An interesting application of IoT in construction is a prototype that was designed for an IoT-based construction site safety management system [117] that can be operated at a low cost independently of the size of the construction site. Additionally, a model for construction site safety monitoring [118] was developed that identifies real-time safety problems and also stores data for future training and improvement. The same paper provided a cost comparison that showed that an IoT system can provide around a 70% cost savings in comparison to traditional systems. Another IoT-based system regarding worker safety with real-time alarming, monitoring, and positioning strategies was introduced in [119], and a safety recognition service using an IoT sensor network in [120]. A commonly reported method is the establishment of a mapping structure between the IoT data and BIM data [76]. The IoT can also be used for integrating environmental and localization data in BIM [121]. Additionally, the IoT can be used in conjunction with the "cloud" to achieve real-time data transmission for monitoring systems, as presented in [122,123]. A similar approach based on the BIM platform for the on-site assembly of prefabricated construction was provided [124], which is enabled by the IoT, while an example of IoT utilization for real-time decision making in repetitive construction operations was presented in [125]. The mentioned applications of the IoT in the construction industry show the versatile benefits it can bring to the construction industry. However, as mentioned earlier, the IoT is a multilayer system that requires wider collaboration for reaching its potential. It is important to research its collaboration with other drivers such as BD.
Big Data (BD)
Due to the rapid development of information and communication technologies, the construction industry is entering the BD era. The term BD can be seen as a rebranding of the term data mining, with a focus on larger and more diverse datasets and sources, with data mining being the technology of discovering structures and patterns in large datasets [126]. BD is believed to be induced with the use of technologies such as radio frequency identification (RFID), and sensor networks [84,127]. Consequently, it is becoming possible to easily collect and effectively use the massive volumes of data that are generated by various design and construction activities to enhance the performance of construction projects [128]. The significance of BD is not to manage a massive amount of data, but to extract valuable information from them.
Referring to the construction industry, BD presents data generated from the life cycle of the building or structures, which includes the phases of planning, design, tendering, construction, checking, and operation management [76]. BD analysis is valuable for more efficient project delivery and for all project stakeholders. Privacy and security issues, skill requirements, data access, sharing of information, and storage and processing issues have been recognized as some of the main challenges of BD [129]. A detailed overview of issues in the field of privacy and security of BD was provided in [130]. Agrawal et al. [131] listed the heterogeneity and incompleteness of data, the scale of data, and the timeliness of analyzing data as the main challenges of BD. Methods of analyzing BD include statistical analysis, online analytical processing (OLAP), and data mining [76]. Besides the expected growth of data from construction business operations, automation of construction processes, safety monitoring/control, resource management, etc. will also lead to significantly more data being generated in the near future [128]. For instance, images from construction activities can be used to identify unsafe behavior of construction workers, with the purpose of reducing the occurrence of safety accidents [132]. Furthermore, multiple studies have reported that BD has the potential to generate immense value for construction projects while effectively improving project performance [84,127]. In addition, based on machine learning, BD can be utilized to accurately predict the performance of construction projects and detect possible uncertainties in project outcomes while still at the early design stage [133]. By mining cost-related data collected from previous projects, strategies can be implemented to control future project costs [84]. An interesting field worth mentioning is the utilization of BD to support smart cities [134]. Lu et al. [135] presented the possibility of using BD in construction waste management and even used BD analytics to identify illegal construction waste dumping [136]. Moreover, it was concluded that an accurate analysis of BD makes it possible to discover new phenomena characteristic of the project, which can help reduce risks in project management [137]. Considering the above, BD can be seen as a crucial component of Construction 4.0 due to the increasing amount of data generated by new IoT devices and BIM-related information.
Additive Manufacturing (AM)/3D Printing
During the past decade, the field of AM, and in particular 3D printing, has gained immense attention in terms of industry usage, technological development, and consumer popularity [138]. It is believed that the first patent for 3D printing dates back to 1984 [139], and experimental applications of AM in the construction industry started appearing in the late 1990s [140]. The paradigm of AM is that a structure can be built by adding an elemental material in a way that can easily be automated [140]. It can also be considered as a process that is based on a three-dimensional digital model that uses automatic technology to create physical objects layer-by-layer without human intervention [141]. If we compare additive with traditional manufacturing, AM offers new possibilities for the design and development of products [49]. It is believed that AM can be a vital component of the 4IR or smart manufacturing due to its high capability as a nontraditional manufacturing approach for mass customization [142]. When talking about construction, AM has the potential to help the construction industry to transition into a technically advanced sector [143], which is proven by the fact that the implementation of AM in the construction industry has resulted in various technical breakthroughs and improvements in construction output efficiency [34].
Perhaps the first case of AM in the construction industry was reported in 1995, when the first construction-scale AM method called "Contour Crafting" (CC) was patented at the University of Southern California [144]. There are various terms encountered in the literature that refer to AM in the construction industry, so printing objects roughly above one cubic meter in volume is referred to as "large-scale AM," or popularly as "large-scale 3D printing" [145]. The International Organization for Standardization and the American Society for Testing and Materials classified AM into seven categories: vat photopolymerization, material jetting, binder jetting, material extrusion, powder bed fusion, sheet lamination, and directed energy deposition [146,147]. However, to date, the processes that are being used for applications in the construction industry include extrusion-based processes and binder jetting [148]. There are various applications of AM in the construction industry and there are even some records claiming different levels of success in printing buildings. For example, two cases of 3D-printed bridges, with one printed from metal and the other from concrete, have been reported [149]. A house in Russia was printed using mobile 3D printing technology, and the building envelope was created in 24 h [150]. In 2015, a five-story apartment building with an area of about 1100 square meters, which is considered the highest 3D-printed structure, was finished [151]. Furthermore, in 2016 the same authors presented the first 3D-printed office in the world [152]. It is suggested that AM could contribute to the construction industry by reducing the exposure of onsite workers to harsh environments and by automating some construction tasks [153]. The main benefits of 3D printing were reported in [149], i.e., new possibilities of design, detailed construction accuracy, reduction of waste, increased safety of workers, possibility of combining different types of materials, and the possibility of printing mechanically connected parts. Additionally, the use of AM in construction could lower the demand for a skilled workforce. On the other hand, imperfections, costs, production duration, limitations of materials, and spatial limitations have been mentioned as the major challenges of AM [142]. The main weaknesses of 3D printing are recognized as possible errors in digital model creation, inappropriate materials, lower production speed, the high price of new technology compared to traditional processes, different mechanical properties caused by material layering, poor surface quality, and the lack of technical standards and regulations [149].
BIM and BD
The essence of BIM can be found in data and information, which explains why BD can be used in the BIM process since BIM will continue to develop. However, a single BIM model is not sufficient to exploit the benefits of BD, and it can take a whole repository of BIM models to be mined for information extracting [154]. Therefore, an emerging trend is the striving for BIM to transfer from personal computers to cloud BIM with the purpose of the project's stakeholders being able to work on BIM from anywhere using their portable devices and access any information necessary for the project [155]. This would enable the data from the cloud sourcing to stay in the BIM and provide support for decision making on a project. Still, the wide usage of BIM is characteristic of the preconstruction stages, while it progressively decreases towards the later stages of a project [156]. Due to the considerable amount of data contained in the BIM model and projects, it can be predicted that BIM could become the center for BD [157,158]. With the accumulation of such amounts of data, the adoption of BIM may be crucial for the creation of a resource for BD analysis [155]. Furthermore, the BD that is contained in BIM can be considered a gold mine for companies to exploit for better decision making and predicting [155]. However, the increasing size and scope of the BIM models are starting to restrict the possibilities of traditional systems being used for storing and processing BIM data [84]. Thus, many cases are going to require customized means of storing and processing BIM. Therefore, BIM-specialized BD storage and processing platforms can be expected [84]. A cloud-based BIM has the potential for providing real-time quantity information due to the advancements of BIM and BD and could stimulate the usage of BIM in the construction phase [155]. Thus far, BIM has been considered to contain only information regarding construction, but the emergence of linked building data is slowly changing this widely accepted attitude [159]. Interesting examples can be found in the connection of BIM with Linked Open Data datasets that contain information on weather, flooding, population density, road congestions, etc. [159,160]. Furthermore, the possibility of using BD in BIM for construction waste minimization was analyzed [161]. A study was conducted in which the need for integrating BIM and BD for maintenance of the lifecycle data and maintenance of the assets and conditions of a highway was stated [162]. Another example was presented in [163], where integration provided cloud computing for the project's members for the facility management. Additionally, a cloud-based system framework was proposed for viewing, analyzing, and storing massive BIM models, and the system was based on Bigtable and MapReduce [164]. The problem with adopting prefabricated construction, i.e., insufficient information for reviewing prefabrication alternatives and choosing suppliers, was recognized in [165], and a system for integrating BIM and BD with the purpose of connecting clients with information about the time and cost of prefabricated elements production was presented. These kinds of integrations of BIM and BD are leading to Big BIM Data, which justifies the emergence of BD as a specialized area of BIM [84]. Additionally, the integration could result in benefits such as better decision making, more efficient modeling and design, failure detection, damage detection, and safety and activity monitoring [84]. Considering the abovementioned applications, the integration of BD with BIM clearly has the potential to reduce the size of data, which can then be used for various purposes in BIM such as weather forecasting, facility management, supplier and alternative selection, and the overall improvement of a construction project.
BIM and IoT
Until recently, the project's stakeholders would enter information regarding their part of the project, i.e., input parameters, in data libraries. This is considered a passive BIM approach [166]. An active BIM approach aims at dynamic data exchange among BIM and sources of input parameters. Active BIM is considered the approach where the integration of BIM and IoT can find application, and a few examples of such integration are presented in the following.
A software architecture for the integration of heterogeneous IoT devices with BIM and GIS was presented. The information from the IoT devices provided the BIM model with actual data and also evaluated the validity of it [167]. A similar example of the integration of environmental and localization data in a cloud-based BIM platform using IoT and BIM was introduced and the platform was validated in two case studies for construction and facility management and operation [121]. Furthermore, an approach using BIM and IoT for construction site management was presented whereby the principle was based on the connection of BIM through a VPL (visual programming language) to a database where information received from the sensors at the site was stored [168]. Such integration could increase productivity and decrease construction duration and costs due to real-time on-site information monitoring. Integration of BIM and the IoT platform for projects involving prefabricated houses was developed. Stakeholders' demands were collected and analyzed and then RFID technology was used for collecting real-time data from the site [124]. Additionally, a system for visual utility tunnel environmental monitoring based on the integration of BIM and IoT was developed [169]. With smart houses becoming the standard, a case study was presented in [170], where the geometry of a building was fundamental in which IoT devices were integrated. It was found that additional software for Autodesk Revit is necessary to be able to visualize and analyze data from the sensors on the 3D model. A paper reviewed all sorts of domains regarding the integration of BIM and IoT, such as construction operation and monitoring, health and safety management, construction logistics and management, facility management, and the methods for its realization [171]. An interesting study was conducted [172] in which the authors integrated BIM and IoT using sensor data and compared it to the model for indoor environment monitoring and comfort analysis. In addition, the system user could judge whether the thermal comfort level had met the standards and the data could guide future equipment choice. Due to the fact that IoT is an emerging technology, and the number of its devices is rapidly growing, the mentioned applications have taken place in the past four years, and it is anticipated that they will continue to develop at an exponential pace.
Discussion
The literature review presented in this research resulted in a total of 172 referenced sources. A standard software tool for constructing and visualizing bibliometric networks, i.e., VOS viewer software, was used to present the keywords referenced in [23]. Figure 2 presents a network map of keywords that appeared in the references based on text data. The keywords were extracted from the abstract and title field of the references applying the full counting method, and with a determined minimum of three occurrences. There were 211 such terms, but the default choice was set to 60% of the most relevant terms, so there were 127 terms analyzed. A further output of the reference network presentation is shown in Figure 3, i.e., mostly referenced journals. This graphical representation supports the relevance of referenced findings, highlighting the journals indexed in Web of Science with a high quantitative bibliometric.
The 4IR differs from the previous industrial revolutions mostly by presenting integrative polyvalent technologies. Therefore, the 4IR already has and will have a much wider scope than the previous industrial revolutions. Even though the 4IR has already taken place in many countries, in developing countries the 3IR has not still fulfilled its potential and it is uncertain when or if the 4IR will appear in those countries. The development of 4IR technologies has stimulated many changes in all sorts of industries, including the construction industry.
Throughout this extensive research review, it is clear that the 4IR is being accepted in the construction industry since the number of papers and the number of presented applications of 4IR technologies in the construction industry is continuously growing. Consequently, the drivers of the 4IR are found to be the drivers of Construction 4.0, except for BIM, which is dominant for the construction industry itself. Although induced by the 4IR, Construction 4.0 is conceptually more focused. An extensive research on the drivers of Construction 4.0 was undertaken. The most important are presented in this paper, with their benefits, challenges, and possibilities of application. The main drivers of Construction 4.0 were found to be BIM, BD, and IoT since they are the keywords that appeared most often in the analyzed papers, as confirmed by the visual representation in Figure 4. These technologies have provided the basis for most of the current advances in the construction industry. The 4IR differs from the previous industrial revolutions mostly by presenting integrative polyvalent technologies. Therefore, the 4IR already has and will have a much wider scope than the previous industrial revolutions. Even though the 4IR has already taken place in many countries, in developing countries the 3IR has not still fulfilled its The 4IR differs from the previous industrial revolutions mostly by presenting integrative polyvalent technologies. Therefore, the 4IR already has and will have a much wider scope than the previous industrial revolutions. Even though the 4IR has already taken place in many countries, in developing countries the 3IR has not still fulfilled its 4.0 was undertaken. The most important are presented in this paper, with their benefits, challenges, and possibilities of application. The main drivers of Construction 4.0 were found to be BIM, BD, and IoT since they are the keywords that appeared most often in the analyzed papers, as confirmed by the visual representation in Figure 4. These technologies have provided the basis for most of the current advances in the construction industry. Using VOS viewer software, a map based on text data was created to give insight into the main Construction 4.0 drivers. Figure 4 is related to Figure 2, presented in the section on research methods, but with an increased number of term occurrences. The keywords were extracted from the abstract and title field with a full counting method, and the minimum number of occurrences of a term was five. There were 211 such terms, but the default choice was set to 60% of the most relevant terms, which resulted in 127 analyzed terms. After determining the main drivers of Construction 4.0, their synergy with BIM was analyzed in order to determine whether it will enable their-but, most importantly, BIM's-full potential in all phases of the construction industry. Despite the potential that Construction 4.0 shows throughout its technologies and application to transform the construction industry, most of these technologies are still mostly represented in the preconstruction phase of the project's lifecycle, in which their use is continually developing. Consequently, this is also the case for BIM, which is used worldwide in the design phase, but Using VOS viewer software, a map based on text data was created to give insight into the main Construction 4.0 drivers. Figure 4 is related to Figure 2, presented in the section on research methods, but with an increased number of term occurrences. The keywords were extracted from the abstract and title field with a full counting method, and the minimum number of occurrences of a term was five. There were 211 such terms, but the default choice was set to 60% of the most relevant terms, which resulted in 127 analyzed terms. After determining the main drivers of Construction 4.0, their synergy with BIM was analyzed in order to determine whether it will enable their-but, most importantly, BIM's-full potential in all phases of the construction industry. Despite the potential that Construction 4.0 shows throughout its technologies and application to transform the construction industry, most of these technologies are still mostly represented in the preconstruction phase of the project's lifecycle, in which their use is continually developing. Consequently, this is also the case for BIM, which is used worldwide in the design phase, but still lacks application in the later phases of a project's lifecycle, from which the construction industry and the project's performances could undoubtedly benefit. The reason behind this can be found in the well-known traditionalism and resistance of the construction industry to changes and innovations, which are mainly caused by traditional practices and the lack of a skilled workforce. The ISO 19650, namely parts 3 and 5 (i.e., ISO 19650-3 Organisation of information about construction works-Information management using building information modelling Part 3:-Operational phase of assets; ISO 19650-5 Organisation of information about construction works-Information management using building information modelling-Part 5: Specification for security-minded building information modelling, digital built environments and smart asset management) backs up the development of BIM in terms of intelligent information systems with a clear intention for increasing the automation and digitalization of the construction production.
The integration of CPS and DT offers a way for construction project teams to bridge the gap between virtual models and physical construction, creating a cyberphysical production system. In this context, the built environment becomes a rich area for the application of the CPS and DT framework for smart buildings, cities, and infrastructures, colloquially called cyberphysical environments, where the built environment becomes increasingly intelligent and digitally connected.
Conclusions
This paper presents a literature review of the industry pre-4IR, the 4IR itself, Construction 4.0 technologies, its origin and its applications, and the synergy of the main Construction 4.0 drivers, i.e., the synergy of BIM with IoT and BD.
With the aim of answering RQ1, Section 3.2 introduced an explanation of the 4IR and Section 4.2 gave the main Construction 4.0 technologies. It was found that the Construction 4.0 drivers are indeed the drivers of the 4IR, i.e., they originated from the 4IR, except for BIM, which is characteristic for the construction industry. Most of these technologies did exist for themselves, but it was the concept of the 4IR that pushed them into wider application and gave them more popularity by increasing interest in the whole concept.
To answer RQ2, each of the Construction 4.0 technologies was analyzed in Section 4.2 and a visual representation was made (Figure 4). It can be concluded that BIM, BD, and IoT are the most represented, i.e., the most significant Construction 4.0 drivers. Ultimately, it was concluded that BIM, IoT, and BD are the main drivers of Construction 4.0. This provided a basis for answering RQ3, in which the aim was to determine the directions of BIM development with regard to the main Construction 4.0 drivers and whether Construction 4.0 is what will push BIM into wider application. The motivation for this question was found in the fact that BIM is mostly applied in the design phase of construction but lacks application in the later phases of a construction project.
In order to answer RQ3, Section 5 was dedicated to the integration of BIM with IoT and BIM with BD as the main Construction 4.0 drivers. It was found that this integration can contribute to the application of BIM itself and the whole Construction 4.0 concept since BIM is becoming the standard for the construction industry. IoT connects all necessary devices for effective monitoring of all project phases, and BD is a requirement for analyzing the huge amounts of data generated in larger construction projects. Improvements in terms of increased productivity and decreased construction duration and costs are anticipated while increasing safety. The reported synergies of BIM with Internet-linked open datasets resulted in real-time information on weather forecasts, flooding risks, population density, road congestions, waste minimization, facility management, decision making, efficient modeling, failure and damage detection, and construction site monitoring. Integration could improve the performance of real-time monitoring while increasing the quality of the entire construction project by enabling more information regarding all project phases.
This provided a basis for answering RQ4, in which the aim was to determine in which phases of the construction project's lifecycle the benefits of Construction 4.0 are most evident. Unfortunately, the answer to this question is still the design (preconstruction) phase, which is continuously developing and becoming more automated, as can be concluded from all the mentioned applications of Construction 4.0 technologies in the construction industry in this paper, while the construction phase still lags behind in terms of its implementation. It is uncertain whether this will change despite the numerous possibilities that BIM offers since the construction industry is resistant to change and does not easily give up established traditional practices. However, BIM is arguably adopting Construction 4.0 requirements and as such could be recognized as BIM 4.0. | 13,210.4 | 2021-08-04T00:00:00.000 | [
"Engineering",
"Computer Science",
"Environmental Science"
] |
Field Weakening Control of a Separately Excited DC Motor Using Neural Network Optemized by Social Spider Algorithm
This paper presents the speed control of a separately excited DC motor using Neural Network (NN) controller in field weakening region. In armature control, speed controller has been used in outer loop while current controller in inner loop is used. The function of NN is to predict the field current that realizes the field weakening to drive the motor over rated speed. The parameters of NN are optimized by the Social Spider Optimization (SSO) algorithm. The system has been implemented using MATLAB/SIMULINK software. The simulation results show that the proposed method gives a good performance and is feasible to be applied instead of others conventional combined control methods.
Introduction
DC motors are used in many applications and industrial fields, because they can provide a high starting torque.It is also possible to obtain speed control over wide range below and above the rated speed [1].It is well known that the speed of separately excited DC motor (SEDCM) can be achieved either by varying the armature voltage (armature control) or the field current (field control) [2].The speed is directly proportional to the armature voltage and inversely proportional to the field current.In armature control, one can get constant reference speed up to rated over the whole load range [3].However, in the field control, constant reference speed up to 120% -130% rated can be achieved but with loss of the motor developed torque.In this paper, neural network controller has been proposed to operate the DC motor in the field weakening region.This controller has been used to reduce the field current and then increase the speed of motor over rated speed.This paper is organized as follows: a mathematical model of SEDC motor operating in both armature and field weakening regions is given in Section 2; combined control of SEDCM with proposed method has been presented in Section 3; a brief description of neural networks is given in Section 4; in Section 5, a Social Spider Optimization (SSO) algorithm has been developed.Simulation results are given in Section 6, and finally, conclusions are given in Section 7.
Modeling of SEDCM
The quivalent circuit of a separately exited DC-Motor is shown in Figure 1 below.Dependent on the quivalent circuit, the mathmetical equations of the motor are obtained using electromechanical energy conversion and torque balance rules as follows [2] [3]: where a v and f v are the applied terminal voltage to the motor and field voltage, a e is the emf induced in the armature winding, The resistances a R and f R are the armature and field resistances, a L and f L are the ar- mature and field inductances, a i and f i are the armature motor current and field current, af L mutual induc- tance, m K is the motor constant, m ω is the motor speed, e T is the internal torque of the motor, J is the rota- tional inertia of motor, m B is the viscous friction of motor and L T the load torque.
Combined Control of SEDCM
The main advantage of DC motors is simple in the speed control.As the motor angular speed m ω , is directly proportional to armature voltage and inversely proportional to the magnetic flux produced by the field current, adjusting the armature voltage and/or the field current will change the rotor angular speed [4].The separately excited DC motor (SEDM) is usually supplied by two controlled voltage sources, that can be controlled rectifiers or choppers.One source is supplying the motor armature winding with the armature voltage, a v , and the other is supplying the motor field winding by the field voltage f v .The speed of the SEDM can be controlled by controlling either the armature voltage a v or the field voltage f v , or both of them.Usually, the voltage speed control method is used to control the speed of the SEDM below its rated speed.In this method, the field voltage f v is kept constant, while the armature voltage a v is varied.A constant field voltage f v produces a constant field current f I , which in turn, produces a constant magnetic flux in the motor air-gap.Hence, the torque produced on the motor shaft remains constant, as the armature current remains constant.
The motor angular speed m ω and the power on the motor shaft are linearly proportional to the armature vol- tage a v .To run the SEDM above its rated speed, the field weakening method is used.Here, the field current f I is reduced by reducing the field voltage f v .The magnetic flux in the motor air gap will be reduced, causing reduction of the back emf a e .The motor armature current will increase, resulting in increasing the motor speed.As a result of that, the back emf a e will increase and a new equilibrium will be established at a higher speed.In other words, with decreasing the field voltage f v , the motor angular speed m ω increases, while the motor torque decreases and the motor shaft power remains constant [5].The block diagram of conventional combined control is shown in Figure 2 below.
In this paper we propose a new method to increase the speed of the motor above the rated value (N rated).The new method based on predict the duty ratio (D) that give the field voltage f v and thus give the require current field f i .This can be realized by using the ability of neural networks (NNs) in prediction and control.Assume that this method don't need to sensors for current field f i and armature voltage a e .Also in this method, the predicted field current will keep the armature voltage of the motor under the rated value.
Neural Network
The structure of an artificial neuron is inspired by the concept of a biological neuron.Neural Network (NN) basically performs input output mapping which can be static or dynamic.One important feature of NN is that it normally requires supervised training (or learning) by input-output example data sets unlike conventional programming of digital computer [6].In learning process, neural network adjusts its structure such that it is able to output the same signals as the supervisor.The learning is repeated until the difference between network output and supervisor is enough [7].The basic model of a single artificial neuron consists of a weighted summer and an activation (or transfer) function as shown in Figure 4.The weights sum j s is therefore ( ) The sigmoid activation function is popular for neural network application.The equation for a sigmoid function is: Feed forward network is a network of single neurons jointed together by synaptic connections.Figure 5 show a three-layer feed forward neural network [8].
The Social Spider Optimization (SSO) Algorithm
A new type of evolutionary technique, which is called Swarm intelligence has enticed much research interest [9].The swarm expression is employed in a general manner to refer to any collection of interactive agents.Swarm intelligence is concerned with the methodology to model the attitude of social animals and insects for problem solving.Researchers invented optimization algorithms by simulating the behavior of ants, bees, bacteria, fireflies and other organisms [10].Self-organization and job division are the basic components of swarm intelligence.Self-organizing system means; each of the covered units responds to local exciter individually and may act together to achieve a global mission, via a job separation which averts a localized supervision.This will efficiently adapt the entire system to internal and external changes [11].A swarm algorithms are built up on several methods.Such methods involve the social behavior of bird herding such as the Particle Swarm Optimization (PSO) algorithm [12], the cooperative demeanor of bee colonies such as the Artificial Bee Colony (ABC) technique [13], the intermarriage demeanor of firefly insects such as the Firefly (FF) method [14] and imitation the cuckoo birds lifestyle such as the Cuckoo Optimization Algorithm (COA) [15].
The SSO supposes that entire search space is a sectarian web, where all the social-spiders react to each other.Each solution within the search space symbolizes a spider position in the communal web.Every spider receives a weight relating to the fitness value of the solution that is denoted by the social-spider.The algorithm supposes two different search agents (spiders): males and females.Depending on gender, each person is behaved by a set of different evolutionary operators which mimic different mutual behaviors that are commonly assumed within the colony.The computational steps for the SSO algorithm can be abstracted as follows [11]: 1) Considering N as the total number of n-dimensional colony members, define the number of male Nm and females N f spiders in the entire population S.
3) Calculate the weight of every spider of S as follows: where ( ) i J s is the fitness value obtained by the evaluation of the spider position i s , { } ( )
4) Move female spider according to the female cooperative operator
The sectarian web is used as technique to transmit information among the colony individuals, this information is encoded as small vibrations.Each vibration depends on the weight and distance of the spider which has generated it.
There are two type of vibration: a) i Vibc which is understood by the member i as response of the information transmitted by the individual c( c s ), where c meant the nearest member to i and assigns higher weight in comparison to i where , Vibb which is understood by the member i as response of the information transmitted by the individual b( b s ), where b meant the best fitness value such that where , The movement of the female is depended on the attraction and repulsion, and they are depended on a uniform random number m r within range [0, 1].If m r is smaller than a threshold PF an attraction is produced, oth- erwise a repulsion is generated.
The new position 1 with probability 2 1 with probability 1 2 where , , and rand α β δ are random value between [0, 1], while k represents the iteration number.5) move the male according to the cooperative operator We have now the third type of vibration i Vibc (a vibration understood by the member i as a result of the in- formation transmitted by the member f( b s ) where f is the nearest female to individual i.
2 , The male individuals are arranged in decreasing order according to their weight value, the individual whose weight f N m w + is located in the middle is considered the median male individual.The change of positions will be as follows: ( ) where f s represents the nearest female to the male i. 6) mating is performed by predominant males and females members, when a dominant male m locates at a set of females within a specific range r, it mates forming a new brood.
r is calculated as follows: ( ) 7) it stop condition is verified the algorithm is finished otherwise go back to step 3.The flow chart of SSO is shown in Figure 6.
Simulation Results
The SIMULINK model of Field Weakening Control for separately excited DC motor have been implemented using MATLAB/SIMULINK software as shown in Figure 7.The ratings and parameters of DC motor are shown in Table 1.PI controller in outer loop have been used to control the speed of the motor in both armature and field modes.To control the current of the motor, Hysteresis current method have been used in inner loop.The maximum current ( a i max) can the motor reach to it is twice the rated current.The voltage of field can be vraied by changing the duty ratio.For the speed equal or less than the rated speed the field current keep constant.For the speed higher than the rated speed the field current is varied to control the speed.Pre-data sets have been used as input and output to NN to training the weights and bais based on SSO.The training data are shown in Table 2.Each weights of the neural network will represent the male and female individuals.The number of individual is assumed to be 60, while the range of parameters is supposed to be between [0, 1].The summation of the mean square error is taken as a fitness function.The structures of NN have three neorons in hidden layer and one neoron for the input layer and another one in the output layer.
The speed and field current responses of the motor under and over the rated speed is shown in Figure 8.The field current is 1.6 A for all speeds under rated speed of the motor which is 1750 r.p.m.A three step changes of 200 r.p.m in a reference speed is applied at t = 12, 20 and 28 s respectively.As shown in the responses (a) and (b), when the speed of the motor N increases over the rated speed Nb the field current decreases.Figure 9 show the summation of the mean square error to the number of training iterations.To drive the motor above rated speed by using field weakening method, the armature voltage should kept constant under the rated value.This can be clearly observed in Figure 10.Through step changes in the motor speed, the armature voltage will remaining almost constant below the rated voltage of the armature which equal to 225 v.
Conclusion
The armature and field circuits of SEDCM are providing from separate sources; this can give a flexible control and wide speed rang.Field weakening methods are dependent on the reverse relationship between the field current and the speed of the motor.To control the speed of the motor from zero up to rated speed, armature control method is used and thus the motor drives in the constant torque region.While controlling the speed over the rated speed (the motor operates in the constant power region), field weakening methods should be used.In this work, the requirement field current has been predicted by using NN to drive the motor in both torque and power regions without needing any sensor to the armature voltage or field current.The parameters of NN optimized using SSO.The simulation results show the effectiveness of the proposed method.
Figure 1 .
Figure 1.The quivalent circuit of a separately exited DC-motor.
Figure 3
below show the block diagram of the proposed method.
Figure 2 .
Figure 2. Block diagram of conventional combined control for SEDCM.
Figure 3 .
Figure 3. Block diagram of proposed field weakening control.
Figure 4 .
Figure 4. Basic model of a single artificial neuron.
Figure 9 .
Figure 9.The summation of the mean square error.
Figure 10 .
Figure 10.Armature voltage of the motor.
Table 1 .
DC motor ratings and parameters.
a L = 0.008 H a
Table 2 .
Data sets for combined control. | 3,438.4 | 2016-01-19T00:00:00.000 | [
"Computer Science",
"Engineering",
"Physics"
] |
5G-Enabled Tactile Internet for smart cities: vision, recent developments, and challenges
The Tactile Internet (TI) is an emerging technology next to the Internet of Things (IoT). It is a revolution to develop smart cities, communities, and cultures in the future. This technology will allow the real-time interaction between humans and machines as well as machine-to-machine with the 1ms challenge to achieve in round trip latency. The term TI is defined by the International Telecommunication Union (ITU) in August 2014. The TI provides a fast, reliable, secure and available internet network that is the requirements of the smart cities in 5G. Tactile internet can develop the part of the world where the machines are strong, and humans are weak. It increases the power of machines so that the value of human power will increase automatically. In this framework, we have presented the idea of tactile internet for the next generation of smart cities. This research will provide a high-performance reliable framework for the internet of smart devices to communicate with each other in a real-time (1ms round trip) using IEEE 1918.1 standard. The objective of this research is expected to bring a new dimension in the research of smart cities.
Because of these features, the internet wants to move from mobile to Tactile Internet. The proposed research work is an enhancement and implementation of a reliable framework based on Tactile Internet emerging technology next to the internet of things [8], [9]. The research outcome is to establish a new reliable framework for communication among humans to machine and machine to machine in the future smart cities [10]. The proposed research uses the correct and efficient simulation of the desired study and can be implemented in a framework of smart cities. In the future, researchers can enhance this research and implement it on the internet of everything framework [11]. Figure 1 shows the applications of 5G.
Fig. 1. 5G Applications
The objective of this research is to create a new reliable communication framework for smart cities using the Tactile Internet the next revolution of the internet of things. This research is based on low-latency, ultra-high availability and high-performance concepts of Tactile Internet. The framework provides QoS through reducing the latency (1ms in the round trip) also the variety of the number of smart devices. In this research, I consider the idle state in order to make our examination more efficient, at that point the general execution regarding the overall performance of the framework is evaluated. The framework will monitor and analyze the real-time data collected from the network and then taking the action.
The research is primarily focused on the next-generation Internet for smart cities. It enables smart devices to communicate with another device among the internet of smart devices using fast, reliable and secure tactile internet. The proposed framework for communication will access the internet of smart devices. The results of the proposed research will be compared with the previous study in the same area.
The rest of the paper is organized as follows. Section II represents the related works, Section III represents the research methodologies, Section IV represents the applications of the Tactile Internet, Section V represents the recent developments, Section VI shows the challenges and Section VII represents the results and discussion and section VIII shows the conclusion.
II. Related Works
In 1991, Theodore S. Rappaport published an article entitles "The wireless revolution", in this paper he presented the wireless communications is the emerging technology as a key for communication among humans as well as devices [12]. In the med of 2006, Amazon achieved a prominent milestone by testing elastic computing cloud (EC 2) which initialized the spark of cloud computing in it. However, the term cloud computing has not coned until March 2007 [13]. The following year brought even more rapid development of the newly emerged paradigm.
Furthermore, cloud computing infrastructure services have widened to include (SaaS) software as a service [14]. In the mid of 2012, the oracle cloud has been introduced, where it supports different deployment models. It is provisioned as the first unified collection of its solutions which is under continues developments. Nowadays, typing cloud computing in any search engine will result in a tremendous result. For example, it would result in more than 139,000,000 matches on Google. [15], In this paper, the authors have presented the tactile internet for smart communities in 5G. They summarize the use of non-orthogonal multiple access protocols in 5G. In the technical report [16], the authors represented the tactile internet as the next revolution after the Internet of things.
In the article [17], the cloud-based queuing model is explained for the tactile internet. In the article [18], the author discussed the haptic communication system. In the article [19], the authors have enabled the tactile internet for ultra-reliability and fast response time (<1ms).
In the paper [20], the authors present a review on tactile internet for industries. It represents the role of tactile internet in the future industries. In the thesis [21], the author explored the challenges and standards for the tactile internet in 5G.
In [22], the 5G-based Tactile internet framework is designed. Very few articles are written on Tactile internet. The previous studies showed us the role, use of tactile internet in the 5 th generation.
III. Research Methodology
The smart devices are increasing exponentially day by day in the whole world [23]. They provide many more facilities to the end-users and also attach to their daily life [24]. Smart devices can connect to the internet easily for sending and receiving data within the network [25].
The smart devices are not just smartphones [26], it may be the smart refrigerator, Smart home automation entry point, smart air conditioners, smart hubs, Smart thermostat, Color-changing smart LEDs, Smart Watches and Smart Tablets, etc. in the internet of things framework, they are connected to each other through the internet [27], [28], [29]. Figure 2 represents the human reaction time. The proposed research plan builds research on extending the performance of communication in the internet of things using tactile internet. The transfer data from one configuration to another using a wireless network starts from 1973 in the form of the packets radio network. They were able to communicate with other same configuration devices. Recent work is continuing a project called the Serval Project. It provides networks facility to android devices for communication in an infrastructure-less network [30], [31]. Whereas our research is concerned about high-performance communication on the internet of smart devices for smart cities.
The main contribution of this research is the creation of a reliable communication framework and provide secure, reliable and fast communication using Tactile Internet among the internet of smart devices. The previous studies have been focused on the creation and optimization of the framework for communication, but such research doesn't perform the full framework for secure and reliable communication among the internet of smart devices for smart cities. Figure 3 shows the tactile internet in 5G.
IV. Tactile Internet Framework
The Tactile Internet necessitates the optimum turnaround time, accessibility, reliability, and security. Such goals can also be accomplished through the construction of a distributed service model. There is a need for ultra-low end-to-end latency to encourage tactile apps to remain regional, near to subscribers. Figure 4 represents the tactile internet working process.
V. Applications
Numerous apps in several diverse areas might be produced through the tactile Internet, such as the following.
A. Health apps
The Tactile Internet is predicted to facilitate public health and safety via the implementation of new apps and services that do not endorse conventional systems. Such apps involve remote rehabilitation, remote treatment, and remote surgical intervention. Such apps would make premium medical training and healthcare professionals accessible everywhere and break the restrictions of place and time [32].
B. Vehicle apps
The Tactile Internet is supposed to support in the management of road transport via traffic detectors as well as driver-assist schemes. The Tactile Internet would enable and to provide the vehicle-to-vehicle connectivity function as well as the vehicle-to-vehicle highway communication networks [32].
C. Apps for industrial automation System
The closed-loop touch with the reality circuit for most industrial automated machines necessitates an end-to-end latency per sensor of 1 ms. It could be accomplished by implies of a tactile Internet structure, as well as the present connected model could be transformed into wireless or enhanced infrastructures. It will allow multiple automated systems or improve the efficiency of the current approach [32].
D. Smart Grid Apps
An intelligent network is built to distribute energy efficiently generated and with the necessary reliability of the power consumption. Smart grid networks effectively control the activity of both power generators and power grids. Also, user usage or tariffs are controlled and monitored by smart grid technologies. These systems, therefore, need ultra-reliability and low-latency communications networks (e.g. the recommended end-to-end latency of the symmetric co-phase of energy supplies is 1 ms) to transmit information over the Internet. The Tactile Internet would defend such structures by accomplishing an end-to-end latency of 1 ms with ultra-high reliability [32].
E. Other Apps
It is expected that the Tactile Internet will have apps in many other important areas such as education, community, professional gameplay, and surveillance drones, etc. The Tactile Internet would help to develop apps that could support children and individuals with disabilities in enhancing their learning skills, recovering from injuries or disability skill sets [32].
VI. Recent Developments
The following are the recent developments in the area of Tactile Internet.
A. Deep Learning-based reliable connectivity
The Tactile Internet requires ultra-reliable connectivity through huge IoT gadgets. The grantfree non-orthogonal multi-access (NOMA) utilizes the combined advantage of grant-free access and non-orthogonal communications to obtain lower latency significant accessibility. Moreover, this suffered from lowered accuracy due to random intrusion. They formulated a variation optimization challenge to enhance the performance of the grant-free access. The deep learning was used to parameterize unsolvable variational functionality with a designed deep neural network [33].
B. Tactile Robots
This new frontier of an interactive coexistence between humans and robots constructs on various new technologies in automation, multi-modal teleoperation, smartwatches, distributed computation, and mobile technologies [34].
C. Tactile Internet Architecture for Smart City
In order to completely integrate tactile technology with smart cities, the new QoE-driven Tactile Internet framework for smart cities including five layers: sensing layer, transmission layer, processing layer, computing layer, and application layer was designed. Specifically, the strategies represented in each layer of this architectural design comply with the criteria of low latency, high reliability, and high user expertise. Within this framework, they are designing a fast and reliable QoE management platform focused on a large training system [35].
VII. Challenges
Clearly, allowing a 1-ms round-trip latency is a major challenge in Tactile Internet. Physical transmission might have very small packets to allow 100 μs of one-way physical layer transmission [36]. This will accomplish with each packet could not surpass the period of the 33-μs packet. A possible explanation for this is the structural extra latency that requires to be implemented by encrypting the packet on the transmitter and detecting and decryption it on the transmitter. It restricts the packet size to less than one-third of the targeted latency. It clearly shows that the synchronization used throughout recent LTE wireless networks is not a feasible solution [37]. Figure 5 shows the roadmap of the wireless network from 1995 to 2030.
VIII. Discussion
By moving to 1 ms round-trip latency together with carrier-grade robustness and accessibility, a new innovation that enables unparalleled mobile apps became feasible. Such systems are called the Tactile Internet as this is the standard latency communication needed for tactile operation and control of real and virtual objects without the development of cybersickness. It will revolutionize education, accessibility and traffic, healthcare, sports, culture, gaming, and the smart grid, just to mention a few of the segments that can be used. The Tactile Internet must radically reshape our culture.
IX. Conclusion
The main contribution of this research is designing a framework for ultra-reliable, low latency and high availability communication on the Internet of smart devices for future smart cities using the Tactile Internet. The proposed framework is specifically appropriate for applications in which data is periodically transmitted on the internet of smart devices environment. In these applications, on one hand, packets are being produced based on a certain period of time pattern. On the other hand, the service time is always a random variable with the general distribution. Therefore, service time might temporarily exceed the period time which, as an inevitable consequence some packets might encounter a busy channel and be dropped. We solve this problem by proposing the new communication framework. We demonstrate that the proposed reliable framework, not only increases the throughput but also the direct connection between the generation (sensors) and communication packet systems are eliminated which makes the system far more stable. Moreover, in order to enhance the proposed model, we have employed a retransmission scheme, variable packet length, and saturated traffic condition. The solution to this research is summarized as follows. The implementation of the proposed framework for communication among the internet of smart devices in 5G will be programmed to execute on to the internet of things using Tactile Internet concepts. The idea will focus on three main concepts, these concepts are Reliability, Security, and availability. The proposed study supports wireless networking technology to establish a reliable framework among the internet of devices for smart cities. | 3,129 | 2019-07-03T00:00:00.000 | [
"Engineering",
"Environmental Science",
"Computer Science"
] |
Identifying faults in the building system based on model predic- tion and residuum analysis
The energy efficiency of the building HVAC systems can be improved when faults in the running system are known. To this day, there are no cost-efficient, automatic methods that detect faults of the building HVAC systems to a satisfactory degree. This study induces a new method for fault detection that can replace a graphical, user-subjective evaluation of a building data measured on site with an automatic, data-based approach. This method can be a step towards cost-effective monitoring. For this research, the data from a detailed simulation of a residential case study house was used to compare a faultless operation of a building with a faulty operation. We argue that one can detect faults by analysing the properties of residuals of the prediction to the actual data. A machine learning model and an ARX model predict the building operation, and the method employs various statistical tests such as the Sign Test, the Turning Point Test, the Box-Pierce Test and the Bartels-Rank Test. The results show that the amount of data, the type and density of system faults significantly affect the accuracy of the prediction of faults. It became apparent that the challenge is to find a decision rule for the best combination of statistical tests on residuals to predict a fault.
Introduction
In the effort to fight global warming, one of the goals of the german government is to achieve a climate-neutral building stock by 2050. The policies focus on two strategies, the use of renewable energies and the increase in energy efficiency. Long-term climate neutrality in the building sector can be achieved by reducing energy consumption and expanding the use of renewable energy [1]. The thermal properties of the building envelope, the efficiency of building technology, and the user behaviour significantly influence energy efficiency of a building. In order to take measures to improve existing buildings, it is important to detect the actual energy efficiency of a building. With the use of on-site measurement data, the actual energy consumption can be detected and flaws in energy efficiency identified.
This study focuses on the energy efficiency of HVAC installations. There are two methods for fault detection, measurement can be performed on-site and faults detected afterwards by analysing the data, or the faults are detected in real-time during operation and reported directly to the building technology manager. Fault detection can be performed by comparing the data collected on-site to a model that delivers expected values. The method is trained with simulation data and applied afterwards to data measured on-site.
Within IEA ECB Annex 71 [2] (International Energy Agency's Energy in Buildings and Communities Program; Annex 71 Building Energy Performance Assessment Based on In-situ Measurements) the members of the Annex were to solve an exercise to explore system identification tech-* Corresponding author<EMAIL_ADDRESS>niques. Compared to the exercise, this study focuses on the faults of the building system. Simulation data of the twin houses, which are two identical case study houses at the Fraunhofer Institute for Building Physics in Holzkirchen, Germany were used. The simulation was carried out within a validation study of the IEA ECB Annex 58 and 71 project [3], [2]. The data consist of two sets, a first part in which fault free operations are simulated and a second part in which various system faults were integrated into the simulation. Both data sets have the length of one month.
In a first phase two different statistical models, a random forest (a machine learning method) and an ARX model (a time series method), predict the normal operation by predicting the total heating power of the building. In the second phase, these two prediction models predict the data set that contains faults. The fault detection is carried out using residual analysis, model checking based on residual analysis is a standard technique for time series analysis, cf. [4], page 175 ff. and [5], page 360 ff. With a suitable time series model adaptation, the residuals are generally assumed to be an approximate white noise and i.i.d. with a mean value of zero. In this study we use this characteristic as the starting point for a decision technique. We propose that the decision method is appropriate for general residuals based on a good model fitting (e.g. resulting from a random forest model).
Fassois and Sakellariou [6] give an overview of time series methods for fault detection in vibrating structures. There are two main types of time series methods for fault detection: non-parametric methods, which use spectral analysis, and parametric methods, which can be categorised as parameter-based and residual-based methods. For the residual technique, the estimations of the model parameters do not have to be considered. The residuals can be calculated directly from the predictions (based on the same modelling method) and the responses. The white noise property of the residuals can be analysed using various statistical tests, which work partially as portmanteau tests. In this study a data-driven decision rule for fault detection merges multiple tests. The approach of combining tests involves a general applicability of the decision rule. Different faults and different prediction methods for the response yield different deviations of the standard behaviour of the resulting residuals. The decision rule for fault is learned on the residual data and can be seen as a sort of portmanteau decision rule for fault detection where the null hypothesis is specified by faultlessness. The technique is adjusted to the situations of observed and unobserved faults in the learning sample. Furthermore, the method is formed for fault detection in specific time points and time intervals.
The developed methods for fault detection could replace a graphical, user-subjective valuation of a residual plot using an automatic, data-based approach.
Description of simulated data and system faults
The simulation was build upon an empirical validation experiment of Annex 58 and Annex 71. A detailed description of the two identical full-size buildings of the Fraunhofer Institute for Building Physics in Holzkirchen, Germany, and data of the Annex 58 experiment can be found in [7], [8], [9] and [10]. The data set is obtained through detailed simulation with the program IDA ICE [11]. The simulation uses the house description of Annex 58 and 71 and the climate boundary conditions of January and February 2019 in Holzkirchen (Annex 71). For the simulation model, each room was equipped with a 2000 W electric radiator with a longwave radiation fraction of 40%. The heating set point of the air temperature was set to 21 • C and controlled room wise by a thermostatic control with a dead band of ±0.5 K. The simulation integrates a MHVR (Mechanical Ventilation with Heat Recovery) air handling unit with a heat recovery of 80 percent with an integrated MVHR summer bypass (possibility to switch off the heat recovery during summer months). The simulation includes a simple occupancy plan of a four-person household with the absence of users between 7:30 and 17:00 each day. The data set starts on January 1st and ends February 28th. The first month (1/Jan -31/Jan), the building runs in regular operation. The second month (1/Feb 28/Feb) includes faults in the operation of the building. In this study three faults are selected. The first fault is a circuit breaker failure (F1; Circuit breaker failure of the electrical heating in the upper floor; 2. Feb. 0:00 -4. Feb. 23:59), the second fault a failure of the MHVRs heat recovery unit (F2; MHVR summer bypass switch off the heat recovery during fault duration; 10. Feb. 0:00 -15. Feb. 23:59), and the third fault a higher thermostat set point than necessary (F3 -Living room thermostat to set temperature 28 • C; 20. Feb. 0:00 -23. Feb. 23:59 ). The data set contains the following indoor and outdoor properties. Air temperature for each room and total heating power supplied by all electrical radiators are measured indoors. The outdoor properties are the air temperature, relative humidity, diffuse and direct solar irradiation on horizontal surfaces, and wind speed and wind direction.
Statistical tests
The presented statistical tests are implemented with the programming language R [12], and the user interface RStudio [13]. The graphics were created with the R package "ggplot2" [14].
The predictive models for the response total heating power use as predictors the indoor temperatures of all rooms, all outdoor information and the daytime in hours. The time-dynamic models take the differences between the value of response at the current time and the value at the time one hour ago (time lag one hour) into account. The total heating power is the model output and the values of the total heating power one hour and two hours ago at each point in time are added as predictors (features) to the models. The predictive modelling is carried out with the method random forest [15] and an ARX (autoregressive with exogenous variables) time series model [16]. To predict the total heating powerŷ t in February, the January data is used for training and the February data for testing. The total heating power in January is predicted with 4-fold crossvalidation. The 4-fold cross-validation divides the January data into four nearly equally sized parts. Then three of these parts predict the reminding part. The difference between the real values (responses) and the predicted values of the total heating power are the residuals. Let y t be the response of the observed total heating power andŷ t the predicted total heating power from a model at the time t = 1, ..., n. Thenε t := y t −ŷ t denotes the residual at the time t andε := (ε 1 , ...,ε n ) the vector of the residuals. Furthermorẽ ϵ := (ε 1 , ...,ε n ,ε n+1 , ...,ε n+L−1 ) := (ε 1 , ...,ε n ,ε 1 , ...,ε L−1 ) is defined for a fixed L ∈ {2, ..., n}. Figure 1 shows the January and the February residuals for the developed random forest and the ARX model.
After successful modelling, the typical properties of residuals are to be statistically tested. A fault in the data process is assumed if the behaviour of the residuals deviates significantly from the standard properties. The special properties of residuals depend on the modelling methodology, the data structure of the learning sample, as well as on the prediction quality of the model. It is assumed that the residuals have a median of zero, are independent and therefore uncorrelated from each other, and behave randomly. The Sign Test [17] and the Wilcoxon Signed-Rank Test [18] are suitable for testing for median equal to zero. The Turning Point Test [19] is well suited for testing independence, the Box-Pierce Test and the Ljung-Box Test [20] for autocorrelation. Randomness can be tested with the Bartels-Rank Test, Cox-Stuart Trend Test, Difference-Sign Test and Mann-Kendall Rank Test [21]. In total, this study examines the residuals using nine tests divided into four test objectives.
Moving p-value
The moving residuals for the shift s with time window length L ∈ {2, ..., n}, which represents the sample size of the moving residuals, are defined by Mε(s, L) := (ε 1+s ,ε 2+s , ...,ε L+s ). The moving residuals are used in order to avoid testing all residuals at once. If p T (.) is the p-value of the statistical test T , it is used (p T (Mε(0, L)), p T (Mε(1, L)), ..., p T (Mε(n − 1, L))), for a fixed L, to examine periods for faults. If the p-value of a test is less then a previously selected significance level α ∈ (0, 1), then the null hypothesis of this test is significantly not met [5], page 5 ff. For all nine tests the null hypothesis is that a certain property of the residuals is fulfilled. Therefore it applies that for each test a fault is suspected when the p-value of this test is smaller then α.
Mean p-value (MPV)
A disadvantage of the p-values of the moving residuals used so far is that it can be recognised at which shift s the p-value is no longer as expected, but not at which time point. In the following, a new constructed function determines faulty time points. This is made possible by a mean of the p-values from the moving residuals. Let Since there are no system faults implemented in the January data a limitation is made to the decision rules which not erroneously detect faults in the January data. Accordingly, a decision rule from the following set is required which is refereed to as choice set.
Parameter optimisations in the case of observed faults
A grid search finds with the known time points of faults an optimised decision rule (L, α, H) ∈ C, which minimises a previously defined fault rate. This optimisation shows on the data of this study good results. The applicability of the model, which was optimised by this procedure on another data, is not tested so far. A problem could be that the decision rule adapts to the data too individually. In future works, the procedure should be repeated with other validation data. Only when data with known faults is available, this technique can be used, which is why future works investigate other techniques for data with unknown faults or without faults.
Results
The January residuals, created by a 4 fold cross-validation, are useful to compare both prediction models. This can be done by using the mean squared error (MSE) of each model, which is defined by 1 n ∑ n t=1ε t 2 . For the ARX model, the calculated MSE equals 31, 001.55, and for the random forest, it is 100, 274.7. For this reason, the ARX model seems to be a better choice to predict the total heating power. We question if a better prediction model is automatically better for fault detection?
Concerning the case that the analysis observed the faults, and the times at which the faults occur are known, the decision rule can be optimised accordingly. This research uses the decision rules of the choice set for each data set and each model with the lowest prediction error rate.
This study analysed whether the ARX model or the random forest model delivers better results when the first, the second or the third fault or all three faults are in the data. Figures 2,3 and 4 show the results when only adding one fault to the data. If only the third fault is in the data the decision rule using the ARX model delivers better result because the decision rule based on the random forest model often suspects faults where no faults are and also matches the existing fault less accurately. If all faults are in the data as shown in Figure 5, both models detect the third fault very well, the ARX model has problems with the second fault, and both models recognise the first fault not well.
Conclusion
This study uses residual analysis for fault detection of HVAC systems in buildings. A detailed simulation of a residential case study house provided the data for the analysis. The predictive modelling of the total heating power was carried out with the method random forest and ARX (autoregressive with exogenous variables) time series model. The residuals were calculated directly from the predictions. A combination of statistical tests explored the white noise properties of the residuals, while a data-driven decision rule that combines multiple tests predicted the faults. The methods for fault detection developed in this study could replace a graphical, user-subjective evaluation of a residual plot using an automatic, data-based approach.
A fault detection method that uses residual analysis has several advantages: For one the method does not depend on the kind of prediction model applied, furthermore information such as model parameter estimations and specific model structures are become superfluous. The research has proved that the methods of fault detection can be applied to a wide variety of data and prediction models such as time series models and procedures of machine learning (including black-box methods). Statistical tests can be added or removed depending on the suspected residual properties.
This study introduces a decision rule when faults are observed. The method depended heavily on the p-values, which usually depend on the sample size, and the methods had to be consistently adjusted to the specific situation and the given sample size. A method that investigates data when faults are unobserved is part of future research. Future studies can avoid the dependency on the p-values by using the test statistics it selves instead of the p-values and defining corresponding threshold values, or by using tests and methods of the Bayesian statistics. Another possible improvement could be to integrate statistical tests that apply the frequency characteristics of the residuals time series.
The method invites the application on data measured on-site. A real building evaluation could be performed using a learning data set based on a building simulation (with simulated, observed faults) of the same building. Consequently, it is essential to test the applicability of a decision rule based on simulated building data on the behaviour of the original building. The accuracy of the building simulation would then be a significant factor in the successful application of the method. Another practical application is the development of a decision rule for fault detection based on real building data set (with enforced, observed faults).
This study used a simulated data set of two months. A future study with a simulated data set with data over one year is planed. In the case that no simulated data are available, but a data set of in situ data, a one month "error-free" data set could be sufficient to determine a significant prediction model. In a further study of Annex 71 [2] in spring 2020, this is the task and the methods developed in this study are to be tested. | 4,159.4 | 2020-09-01T00:00:00.000 | [
"Engineering",
"Environmental Science"
] |
Solutions of Detour Distance Graph Equations
Graph theory is a useful mathematical structure used to model pairwise relations between sensor nodes in wireless sensor networks. Graph equations are nothing but equations in which the unknown factors are graphs. Many problems and results in graph theory can be formulated in terms of graph equations. In this paper, we solved some graph equations of detour two-distance graphs, detour three-distance graphs, detour antipodal graphs involving with the line graphs.
Introduction
The recent rapid growth in the Internet of things has necessitated the development of new approaches to persistent issues in wireless sensor networks. These issues include minimum obstacles in the end-to-end communication path, location accuracy, latency, and delay, among others. These problems can be mitigated by using the distance graph to create local algorithms, i.e., algorithms with minimum communication rounds. We study distance graph applications in wireless sensor networks with a focus on minimum path obstacles and high localization accuracy.
A wireless sensor network (WSN) is a network of tiny wireless sensors that can sense a parameter of interest. The sensed data is forwarded to a base station through the formed ad hoc network of sensor nodes. There are many application areas of WSNs, including M2M communication and the Internet of Things (IoT). WSNs are a fundamental building block of smart homes, smart workplaces, and smart cities, among others. It has a lot of other essential purposes for modern technology, such as scientific research, rescue operations, and scientific discoveries. As sensors are a power constraint tiny devices, energy conservation for extending the network's lifetime is a challenging issue. A wireless sensor network's lifetime heavily depends on innovative schemes that mitigate energy consumption. The distance graph is used to form and localize an ad hoc network of sensor nodes so that sensed data can be forwarded to the base station with little energy cost.
Therefore, finding the solutions to these graph equations is essential. A lot of research has been done by many researchers during the past fifty years, and their results have made a significant contribution in graph theory.
Let X be a nontrivial finite connected graph. Every graph X with detour distance D defines a metric space. The n-distance graph of X, denoted by T n (X), is the graph with V(T n (X)) = V(X) in which two vertices u and v are adjacent in T n (X) if and only if d X (u, v) = n. Furthermore, for each set S ⊆ dist(V(X); D), the detour distance graph Γ D (V(X), S) is the graph having V(X) as its vertex set and two vertices u and v in this graph are adjacent if and only if D(u, v) ∈ S. We denote this graph simply by D (X, S). The graph D n (X) := D (X, {n}) is called the detour n-distance graph of X. If n equals the detour diameter of X, then this graph is called the detour antipodal graph of X, which is denoted by DA (G). A graph X is said to be a detour self n-distance graph if D n (X) ∼ = X.
The line graph L(G) of a graph G is the graph whose vertices correspond to the edges of G and wherein two vertices are adjacent in L(G) if and best if the corresponding edges are adjacent in G.
In this work, we consider the detour distance graphs. Specifically, we solve the graph equations involving detour two-distance graphs, detour three-distance graphs, detour antipodal distance graphs, line graphs, and the complement of graphs. In addition, we solve some equations involving two-distance graphs, three-distance graphs, and antipodal distance graphs with the graphs mentioned earlier.
Harary, Heode, and Kedlacek first studied the two-distance graph. They investigated the connectedness of two-distance graphs. This graph and the relationship between the two-distance graph and line graph was further studied in [1][2][3][4][5][6][7][8][9]. In 2014, Ali Azimi and Mohammad D. X solved the graph equations T 2 (X) ∼ = P n and T 2 (X) ∼ = C n . In 2015, Ramuel P. Ching and Garces gave three characterizations of two-distance graphs and found all the graphs X such that T 2 (X) ∼ = kP 2 or K m ∪ K n , which can be found in [10]. S. K Simic and some mathematicians solved some graph equations of line graphs, which can be found in [11][12][13][14][15][16][17]. In 2017, R. Rajkumar and S. Celine Prabha solved some graph equations of two-distance, three-distance and n-distance graph equations, which can be found in [18][19][20]. In 2018, R. Rajkumar and S. Celine Prabha found the characterization of the distance graph of a path which was described in [21].
Motivated by the results listed above, we solved some graph equations of distance graphs and line graphs. Other graph-theoretic terms and notations that are not explicitly defined here can be found in [2].
Solutions of Graph Equations of Detour Two-Distance Graphs and Line Graphs
First, we consider some graph equations of type D 2 (G) ∼ = G 1 ∪ G 2 , where G 1 and G 2 are given graphs and investigate the solution G of these equations.
The main result of solving graph equations involving detour two-distance graphs we prove in this section is the following: Theorem 1. Let G be a graph. Then, If G is connected, then Lemma 1. Let n ≥ 3 be an integer. Then, Let v be a vertex in G which is not in the cycle of G. Without loss of generality, we may assume that v is adjacent to v 1 . Then, If G is the graph other than these graphs, then by the argument of part (a), By the structure of G, the graph L(G) is connected non-unicyclic. Let the maximum length among all the cycles in L(G) be k. Clearly k ≥ 4. If k ≥ 5, then any vertex in such a cycle is isolated in D 2 (L(G)). If k = 4, then G is the graph obtained from K 3 by adding a pendent edge to any of its vertex. In this case, ) has two isolated vertices. However, G has exactly one isolated vertex. Thus, If G is the graph other than these graphs, then the maximum length among all the cycles in L(G) is at least six. Then, any vertex in such a cycle is isolated in D 2 (L(G)). Therefore, D 2 (L(G)) G.
6.
By part (e), D 2 (L(G)) is disconnected. Thus, by a similar argument as in part (d), we get D 2 (L(G)) G.
If G is the graph other than these graphs, then by the argument of part (a), D 2 (G) is disconnected and so T 2 (D 2 (L(G))) is disconnected. Hence, L(G) is connected. Therefore, T 2 (D 2 (L(G))) L(G).
Combining Propositions 1-3, we get the proof of Theorem 1.
Solutions of Graph Equations of Detour Three-Distance Graphs and Line Graphs
The main result on solving graph equations involving three-distance graphs we prove in this section is the following: Theorem 2. Let G be a graph. Then, If G is connected, then To prove the above theorem, we start with the following: Lemma 2. Let n ≥ 4 be an integer. Then, If n ≥ 7, then the detour distance between any two vertices of C n is at least four. Therefore, D 3 (C n ) ∼ = K n . Proposition 4. Let n ≥ 4 be an integer. Then, 1.
Proposition 5. Let G be a graph.
Proof. 1.
Let G be connected non-unicyclic. Suppose that D 3 (L(G)) ∼ = G. Then, we have |V(G)| = |E(G)|, so G is unicyclic, since G is connected, which is a contradiction to our assumption that G is non-unicyclic. Thus, D 3 (L(G)) G. The proof of the rest of the cases are similar to the above.
Proof. 1.
By the structure of G, the graph L(G) is connected non-unicyclic. Let the maximum length among all the cycles in L(G) be k. Clearly, k ≥ 4. If k ≥ 7, then any vertex in such a cycle is isolated in D 3 (L(G)). If k = 4, 5, 6, then D 3 (L(G)) ∼ = C 4 , C 5 and 3K 2 , respectively. Thus, D 3 (L(G)) is disconnected and D 3 (L(G)) G.
2.
By part (a), D 3 (L(G)) is disconnected. However, G is connected except for the graph C 3 (r, 0, , 0), r ≥ 1. Thus, D 3 (L(G)) G. If G ∼ = C 3 (r, 0, , 0), r ≥ 1, then D 3 (L(G)) has two isolated vertices. However, G has exactly one isolated vertex. Therefore, If G is the graph other than these graphs, then the maximum length among all the cycles in L(G) is at least seven. Then, any vertex in such a cycle is isolated in D 3 (L(G)). Thus, D 3 (L(G)) G.
4.
By part (c), D 3 (L(G)) is disconnected. Therefore, the rest of the proof is similar to part (b).
The proof is similar to the proof of part (b), since D 3 (D 3 (L(G))), T 2 (D 3 (L(G))) and T 3 (D 3 (L(G))) are disconnected. (i)-(j): The proof of part (i) and (j) are similar to the proof of parts (c) and (d), respectively, since D 3 (L(G)) is disconnected.
Combining Propositions 4-6, we get the proof of Theorem 2.
Solutions of Graph Equations of Detour Antipodal Graphs and Line Graphs
The main result on solving graph equations involving detour antipodal graphs we prove in this section is the following: T 2 (DA (L(G))) ∼ = G if and only if G ∼ = C n , where n ≥ 5 and n is odd; 10. T 2 (DA (G)) ∼ = G if and only if G ∼ = C n , where n ≥ 5 and n is odd;
if and only if n = 3, 5.
If G is the graph other than these graphs, then by the argument of part (a), DA (G) is disconnected. Thus, L(G) is connected. Hence, none of D 2 (DA (G)), D 3 (DA (G)), T 2 (DA (G)), T 3 (DA (G)) and A(DA (G)) is isomorphic to L(G).
Combining Propositions 7-9, we get the proof of Theorem 3.
Conclusions
Given a set of wireless sensor nodes and connections, graph theory provides a useful tool to simplify the many moving parts of dynamic systems. In this work, we mainly focused on the study of detour distance graph equations. In particular, we solved some graph equations involving detour two-distance graphs, detour three-distance graphs, detour antipodal graphs and line graphs. This solution is believed to be useful for many researchers and businesses working in wireless sensor networks. | 2,528.4 | 2022-11-01T00:00:00.000 | [
"Mathematics",
"Computer Science",
"Engineering"
] |
Physico-Chemical Characterization and In Vitro Biological Evaluation of a Bionic Hydrogel Based on Hyaluronic Acid and l-Lysine for Medical Applications
Hyaluronic acid (HA) is an endogenous polysaccharide, whose hydrogels have been used in medical applications for decades. Here, we present a technology platform for stabilizing HA with a biocrosslinker, the amino acid l-Lysine, to manufacture bionic hydrogels for regenerative medicine. We synthetized bionic hydrogels with tailored composition with respect to HA concentration and degree of stabilization depending on the envisaged medical use. The structure of the hydrogels was assessed by microscopy and rheology, and the resorption behavior through enzymatic degradation with hyaluronidase. The biological compatibility was evaluated in vitro with human dermal fibroblast cell lines. HA bionic hydrogels stabilized with lysine show a 3D network structure, with a rheological profile that mimics biological matrixes, as a harmless biodegradable substrate for cell proliferation and regeneration and a promising candidate for wound healing and other medical applications.
Introduction
Hyaluronic acid (HA) is a structural building block widely present in the human body throughout the extracellular matrix, and vitreous, connective, epithelial and neural tissues [1,2]. Despite its structural simplicity and repetitiveness, a glycosaminoglycan consisting of N-acetylglucosamine-glucuronic acid, disaccharide units repeating in a linear fashion, HA plays a role in a broad spectrum of physiological processes, such as in morphogenesis and tissue organization, cell proliferation, differentiation and migration, among others [3][4][5].
HA is biocompatible, biodegradable, non-immunogenic and commercially available from non-animal sources with a high purity level at an affordable price, and for these reasons HA medical products have been used for decades in ophthalmology, joint health and facial aesthetics [6][7][8]. The biochemistry of HA is a field of intensive research due to the wide-ranging roles of this multitasking molecule in living systems and for developing further medical applications [3,9].
In living tissues, HA is stabilized by glycoproteins, together with collagen and elastin, building up the extracellular matrix 3D network and giving resilience to connective tissues and supporting cell proliferation, migration and, ultimately, regeneration. These characteristics make HA a promising candidate for applications in tissue engineering, wound healing and soft tissue regeneration [10][11][12][13][14].
In the human body, HA homeostasis is a dynamic process that involves a competition between synthesis and degradation. The exact HA degradation kinetics are dependent on the tissue, but the half-time of degradation normally ranges from hours to days [15]. The hyaluronidase enzymes HYAL 1 and HYAL 2 have been identified as the most active moieties in biodegradation of HA through hydrolysis of β-1,4-glycosidic linkages between N-acetyl-glucosamine (NAG) and D-glucuronic acid [16,17].
Therefore, for specific medical applications, e.g., restoration of age-related volume loss in facial soft tissue, it is important to stabilize HA through a crosslinking process, which turns the otherwise liquid viscous solution into a solid-like biohydrogel, that resembles the natural structure of HA in the body and allows one to tailor mechanical, biological and rheological properties (including injectability) and resorption time [18,19].
Even if BDDE-crosslinked hyaluronic acid hydrogels are considered safe, BDDE itself is toxic and it is mandatory to purify the hydrogel to remove unreacted BDDE (residues must be below 2 ppm). Moreover, BDDE is a synthetic bridge with no potential biological activity to support cell proliferation.
A less common chemical modification of HA is esterification with benzyl alcohol, used to manufacture dressings and films for wound healing and advanced wound care (e.g., products of the HYAFF ® family) [21]. Other chemical routes to modify HA include, for example, the use of glutaraldehyde, divinyl sulfone, adipic acid dihydrazide derivatives, biscarbodiimide, acrylation and oxidation, or activation with N-(3-Dimethylaminopropyl)-N -ethylcarbodiimide hydrochloride (EDC) and N-hydroxysuccinimmide (NHS). However, they all have very limited applications in the market because of safety concerns or because of complex and expensive manufacturing protocols [22][23][24][25][26].
Bionic design consists of integrating information from biological systems into the design and development of new products, in order to develop novel medical tools with good safety and efficacy profiles. We designed a bionic HA hydrogel stabilized with a physiologically occurring, bifunctional biomolecule, L-Lysine, which has the potential to support the repair of injuries or the age-related impaired structures or functions of living tissues.
Indeed, L-Lysine is an amino acid physiologically present in the body and is known for playing an important role in cell adhesion and collagen crosslinking [27,28]. However, pure L-Lysine is not reactive enough to react with the carboxylic or hydroxylic groups of HA. In this work, we present a new, straightforward platform technology for the crosslinking of hyaluronic acid with L-Lysine by using EDC/NHS, non-toxic coupling agents, that allow the preparation of homogeneous bionic hydrogels with a three-dimensional network, with tunable composition, HA concentration, degree of modification and rheological profile, tailored for specific uses in regenerative medicine. Further, we describe and present the results of physico-chemical characterization, degradation kinetics and in vitro biological evaluation.
All other chemicals were of the highest purity available. Phosphate-buffered saline (PBS) solutions were prepared as described:
HA Crosslinking
In a typical synthesis, 1 g of HA (2.6 mmol) is dissolved in 20 mL of a PBS solution at pH = 6. Then, 0.33 g of NHS (2.9 mmol) are dissolved in 1.25 g of water and added to the HA solution, followed by the addition of 1.50 g of EDC hydrochloride (7.8 mmol) dissolved in 2 g of water.
Afterwards, 1.00 g of a 25% solution of L-Lysine (1.7 mmol) in PBS at pH = 7.4 is added and the hydrogel obtained is left to react for 18 h. For the low crosslinked sample HA20L, 0.80 g of lysine solution are employed instead of 1.00 g.
The hydrogel is purified through dialysis for 5 days against PBS at pH = 7.4. The initial (before dialysis) and final net weight are recorded and the HA concentration is adjusted as desired by adding PBS at pH = 7.4 (HA30 = 30 mg/mL; HA25 = 25 mg/mL; HA20 and HA20L = 20 mg/mL).
The hydrogel is autoclaved at 121 • C for 11 min and stored in sterile conditions.
Rheological Analysis
Oscillatory rheological analysis was performed with a DHR3 rheometer (TA Instruments, New Castle, DE, USA) equipped with a 35 mm parallel plate geometry, at a constant temperature of 37 • C to simulate the conditions in the human body.
Compression experiments were performed on a DHR3 rheometer (TA Instruments) equipped with rough 25 mm parallel plate geometry. For these tests, the gap was set to 1 mm. A frequency sweep from 0.5 to 5 Hz at 0.1% strain was performed. The gap was then set to 0.9 mm at 5 µm/s. Another frequency test was performed. The gap was then set to 1 mm again at the same speed. Another frequency test was performed. The cycle was then repeated 10 times in total.
To evaluate cohesivity and stretchability, extensional measures were carried on a Caber rheometer (Thermo Fisher, St. Louis, MA, USA) with 4 mm steel pads. The tests were performed over a distance of 10 mm in 9 s and the evolution of the normalized sample diameter with time was recorded.
Microscopy
Optical images were obtained with a Keyence VHX-600 digital microscope (Osaka, Japan). SEM images were obtained in an environmental scanning electron microscope (FEI XL30 ESEM, FEI Technologies Inc., Fremont, CA, USA). The ESEM investigations were performed in high-vacuum mode of the microscope. The signal was processed with a signal mix of a backscatter electron detector (BSE detector) and a secondary electron detector (SE detector).
The samples were frozen directly in liquid nitrogen and freeze dried for 24 h in a Christ GAMMA 1-16LSG. Afterwards, the samples were mounted on a sample holder and covered with gold.
Enzymatic Degradation Test
The enzymatic degradation of tested hydrogels was evaluated using a previously described protocol for the quantification of the released N-acetyl glucosamine (NAG) [29]. Hydrogels were weighed (0.2 g) and centrifuged in glass tubes at 1000× g using a refrigerated bench centrifuge (Megastar 600R, VWR, Milano, Italy). The hyaluronidase solution was prepared in a specific concentration (6080 U/mL) in isotonic phosphate-NaCl buffer at pH 7.4 and added onto the surface of the gels. After incubation, at different time points (1 h, 3 h, 6 h, 24 h, 48 h, 72 h, 120 h, 168 h), the enzymatic reaction was stopped by the addition of potassium tetraborate solution (0.8 M, pH 9.1), followed by vortexing and heating at 100 • C. NAG released in the solution was assayed according to the methods reported in the literature [30]. Briefly, Ehrlich's reagent (Merck, Darmstadt, Germany) was diluted 1:10 in acetic acid (Merck) and added to the tubes; then, samples were vortexed and incubated for 20 min at 37 • C, to develop a violet color proportional to the NAG content in each sample. After centrifugation at 1000× g for 15 min, absorbance was recorded at a 585 nm wavelength with a microplate reader (Multiskan, Thermo Scientific, Waltham, MA, USA) against a blank prepared with only phosphate buffer and the Ehrlich's reagent.
Data Analysis
Data obtained from hyaluronidase sensitivity tests were analyzed by determining the NAG degradation percentage at each time point. The expected amount of NAG in each sample starting from the percentage of hyaluronic acid in each product was calculated.
The obtained values were used as a reference to calculate the percentage of NAG released by hyaluronidase. The obtained data were plotted using the standard hyperbole equation (GraphPad Prism, San Diego, CA, USA): Data points fitting to the model were evaluated by calculating the R 2 for each sample analysis. Slope values between points 0-50% were calculated as well in order to determine the degradation rate for each product. t 1 2 is defined as the time at which NAG degradation is half of the maximum; t 50% is defined as the time at which NAG degradation is equal to 50%.
In Vitro Cell Biocompatibility
The evaluation of biocompatibility of human fibroblasts after seeding on HA25 hydrogel was assessed by optical microscopy and 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT, Merck, Darmstadt, German) staining. Briefly, hydrogels (0.2 mL) were put inside a Petri dish (µ-Dish 35 mm high, Ibidi, Gräfelfing, Germany) and dermal fibroblasts were seeded on it at a density of 100,000 cells in complete culture medium. Then, after incubation at 37 • C, a morphological observation at different times (t = 0 h, 24 h, 48 h, 72 h and 144 h of incubation) was performed using an optical microscope (VisiScope IT415 PH, VWR part of Avantor, Milan, Italy). Afterwards, viable cells were evaluated using MTT staining. Briefly, after 144 h of cell incubation on hydrogels, a solution of MTT (1 mg/mL) was added at 37 • C for 2 h; samples were visualized under an optical microscope.
Confocal Analysis of Cell Viability
Cell morphology and cell viability were also observed using a confocal microscope (Leica TCS SP8 STED 3X, Wetzlar, Germany). Normal human fibroblasts were seeded (100,000) on HA25 hydrogel, previously placed on a specific support (Ibidi, Gräfelfing, Germany), that led to a uniform distribution of the sample and better growth of cells on it. At different times, staining with the LIVE/DEAD™ Viability/Cytotoxicity Kit (Thermo Fisher Scientific) was performed. Briefly, a solution of calcein AM 2 µM and EthD-1 solution 4 µM was prepared in PBS (Sigma-Aldrich, St. Louis, MO, USA) and added to the samples; after incubation for 30 min, samples were analyzed by confocal microscopy.
Rheological Characterization
The protocol allows for fine tuning of the rheological properties of the final hydrogel by playing with parameters such as molecular weight, initial HA concentration, final HA concentration and HA/EDC/NHS/lysine ratio. For example, Table 1 shows the rheological properties of different hydrogels prepared in different conditions. All the measures were performed after autoclaving the hydrogel at 121 • C for 11 min. In Figure 1, we compare the rheological properties of a 2.5% solution of pristine HA with a 2.5% crosslinked hydrogel (HA25). Strain sweep and frequency sweep confirm that while pure HA mostly behaves as a viscous fluid (G < G ), HA crosslinked with L-Lysine in the same conditions behaves as an elastic gel (G > G ) and no crossover point was observed. This is a clear indication of the formation of a crosslinked network.
Confocal Analysis of Cell Viability
Cell morphology and cell viability were also observed using a confocal microscope (Leica TCS SP8 STED 3X, Wetzlar, Germany). Normal human fibroblasts were seeded (100,000) on HA25 hydrogel, previously placed on a specific support (Ibidi, Gräfelfing, Germany), that led to a uniform distribution of the sample and better growth of cells on it. At different times, staining with the LIVE/DEAD™ Viability/Cytotoxicity Kit (Thermo Fisher Scientific) was performed. Briefly, a solution of calcein AM 2 μM and EthD-1 solution 4 μM was prepared in PBS (Sigma-Aldrich, St. Louis, MO, USA) and added to the samples; after incubation for 30 min, samples were analyzed by confocal microscopy.
Rheological Characterization
The protocol allows for fine tuning of the rheological properties of the final hydrogel by playing with parameters such as molecular weight, initial HA concentration, final HA concentration and HA/EDC/NHS/lysine ratio. For example, Table 1 shows the rheological properties of different hydrogels prepared in different conditions. All the measures were performed after autoclaving the hydrogel at 121 °C for 11 min. In Figure 1, we compare the rheological properties of a 2.5% solution of pristine HA with a 2.5% crosslinked hydrogel (HA25). Strain sweep and frequency sweep confirm that while pure HA mostly behaves as a viscous fluid (G′ < G″), HA crosslinked with L-lysine in the same conditions behaves as an elastic gel (G′ > G″) and no crossover point was observed. This is a clear indication of the formation of a crosslinked network.
We decided to explore the behavior of the hydrogel not only with oscillatory rheology, but also under repetitive axial stress to evaluate the response of the material under the conditions it may experience when used as a dermal filler or in intra-articular injections. We decided to explore the behavior of the hydrogel not only with oscillatory rheology, but also under repetitive axial stress to evaluate the response of the material under the conditions it may experience when used as a dermal filler or in intra-articular injections.
In Figure 2, we show the results of such experiments, in which repeated cycles of compression at different frequencies were performed on an L-Lysine crosslinked HA hydrogel with a 3.0% concentration. The HA-lysine hydrogels are able to withstand compression-elongation cycles at different frequencies without any significant change in their response, and thus no damage to their inner structure. Even at higher frequencies, the differences in the loss and storage modulus are negligible between each cycle.
In Figure 2, we show the results of such experiments, in which repeated cycles of compression at different frequencies were performed on an L-lysine crosslinked HA hydrogel with a 3.0% concentration. The HA-lysine hydrogels are able to withstand compression-elongation cycles at different frequencies without any significant change in their response, and thus no damage to their inner structure. Even at higher frequencies, the differences in the loss and storage modulus are negligible between each cycle. We also decided to perform extensional experiments that simulated the "finger test" commonly used for similar materials in the field of aesthetic medicine. In this test, each hydrogel was placed between the plates of a rheometer and the upper plate was raised in nine seconds by 10 mm. The experiment was performed on three HA-Lys hydrogels and for a standard commercial 28 mg/mL BDDE-crosslinked HA hydrogel as a reference (Renée Volume). Elasticity was calculated by interpolating the experimental data using the following mathematical model:
4
where D0 is the initial diameter of the filament (in meters); G the elastic modulus of the sample (in Pascal) and λc the relaxation time (in seconds). Surface tension was fixed at 60 mN/m and density at 1033 kg/m 3 . In Table 2 we report the elastic modulus of the three hydrogels compared to a commercial dermal filler.
Imaging
We investigated the microscopic morphology of the hydrogel both through standard optical microscopy and through SEM microscopy.
In Figure 3, we show the comparison at a 10× magnification of a classical biphasic hydrogel (Figure 3a) with HA-Lys hydrogel HA30 (Figure 3b), after staining of hyaluronic acid with toluidine blue. It is possible to observe that while classical monophasic HA hydrogels have a non-homogeneous structure, with crosslinked particles dispersed in uncrosslinked or poorly crosslinked gel, HA-Lys hydrogels have an isotropic, homogeneous We also decided to perform extensional experiments that simulated the "finger test" commonly used for similar materials in the field of aesthetic medicine. In this test, each hydrogel was placed between the plates of a rheometer and the upper plate was raised in nine seconds by 10 mm. The experiment was performed on three HA-Lys hydrogels and for a standard commercial 28 mg/mL BDDE-crosslinked HA hydrogel as a reference (Renée Volume). Elasticity was calculated by interpolating the experimental data using the following mathematical model: where D 0 is the initial diameter of the filament (in meters); G the elastic modulus of the sample (in Pascal) and λ c the relaxation time (in seconds). Surface tension was fixed at 60 mN/m and density at 1033 kg/m 3 . In Table 2 we report the elastic modulus of the three hydrogels compared to a commercial dermal filler.
Imaging
We investigated the microscopic morphology of the hydrogel both through standard optical microscopy and through SEM microscopy.
In Figure 3, we show the comparison at a 10× magnification of a classical biphasic hydrogel (Figure 3a) with HA-Lys hydrogel HA30 (Figure 3b), after staining of hyaluronic acid with toluidine blue. It is possible to observe that while classical monophasic HA hydrogels have a non-homogeneous structure, with crosslinked particles dispersed in uncrosslinked or poorly crosslinked gel, HA-Lys hydrogels have an isotropic, homogeneous structure. SEM analysis (Figure 3c,d) confirms that the material is composed of a crosslinked three-dimensional network of interconnected pores, with diameters ranging from 20 to 100 µm. This peculiar microscopic structure is extremely interesting, because it opens the possibility of employing the hydrogel as a scaffold for 3D cell culture, or as an ECM substitute in regenerative medicine. structure. SEM analysis (Figure 3c,d) confirms that the material is composed of a crosslinked three-dimensional network of interconnected pores, with diameters ranging from 20 to 100 μm. This peculiar microscopic structure is extremely interesting, because it opens the possibility of employing the hydrogel as a scaffold for 3D cell culture, or as an ECM substitute in regenerative medicine.
Degradation
The data collected show that hydrogels are sensitive to bovine testes hyaluronidase degradation from the shortest timeframes analyzed. The best-fitting curve was obtained by plotting the absorbance mean of each sample, at each timepoint, with the rectangular hyperbola equation (Figure 4), obtaining the coefficients of determination as an index of data goodness of fit and the degradation parameters reported in Table 3.
Degradation
The data collected show that hydrogels are sensitive to bovine testes hyaluronidase degradation from the shortest timeframes analyzed. The best-fitting curve was obtained by plotting the absorbance mean of each sample, at each timepoint, with the rectangular hyperbola equation (Figure 4), obtaining the coefficients of determination as an index of data goodness of fit and the degradation parameters reported in Table 3.
Degradation
The data collected show that hydrogels are sensitive to bovine testes hyaluronidase degradation from the shortest timeframes analyzed. The best-fitting curve was obtained by plotting the absorbance mean of each sample, at each timepoint, with the rectangular hyperbola equation (Figure 4), obtaining the coefficients of determination as an index of data goodness of fit and the degradation parameters reported in Table 3. The results show that HA20 presents the maximal percentage of NAG degradation (82.30%) and the quickest degradation rate (1.25%/h) with respect to HA30 (NAG degradation of 70.74% and degradation rate of 0.82%/h), demonstrating that the two degradation parameters are not dependent on the initial HA concentration, unlike the T 50% index. Indeed, a higher value of T 50% in HA30 demonstrates that the HA content is mostly associated with the initial hydrogel degradation [29]. As shown in Figure 5, the progressive time-dependent increase in the percentage of NAG release is statistically significant at all times analyzed except for the early ones (1/3/6 h) and the later ones (120/168 h), when it is possible that the release rate reached a plateau. In the case of HA30, the difference from 48/72 h is also not statistically significant. The results show that HA20 presents the maximal percentage of NAG degradation (82.30%) and the quickest degradation rate (1.25%/h) with respect to HA30 (NAG degradation of 70.74% and degradation rate of 0.82%/h), demonstrating that the two degradation parameters are not dependent on the initial HA concentration, unlike the T50% index. Indeed, a higher value of T50% in HA30 demonstrates that the HA content is mostly associated with the initial hydrogel degradation [29]. As shown in Figure 5, the progressive time-dependent increase in the percentage of NAG release is statistically significant at all times analyzed except for the early ones (1/3/6 h) and the later ones (120/168 h), when it is possible that the release rate reached a plateau. In the case of HA30, the difference from 48/72 h is also not statistically significant.
In Vitro Evaluation of Hydrogel Biocompatibility
HA-based products are in direct and prolonged contact with human skin and mucous membranes and, consequently, they should show no or very low toxicity to skin cells or epithelia. In preliminary experiments, to evaluate the biocompatibility of the HA25 hydrogel, a normal human dermal fibroblast (NHDF) cell line was seeded on top of the hydrogel.
Subsequently, a morphological observation was performed using an optical microscope at different times (0/24/48/72 h and 144 h).
As shown in Figure 6, after 24 h, cells have a characteristic elongated shape that is maintained at all analyzed times up to 144 h of incubation with the hydrogels (Figure 7), suggesting the presence of viable cells.
In Vitro Evaluation of Hydrogel Biocompatibility
HA-based products are in direct and prolonged contact with human skin and mucous membranes and, consequently, they should show no or very low toxicity to skin cells or epithelia. In preliminary experiments, to evaluate the biocompatibility of the HA25 hydrogel, a normal human dermal fibroblast (NHDF) cell line was seeded on top of the hydrogel.
Subsequently, a morphological observation was performed using an optical microscope at different times (0/24/48/72 h and 144 h).
As shown in Figure 6, after 24 h, cells have a characteristic elongated shape that is maintained at all analyzed times up to 144 h of incubation with the hydrogels (Figure 7), suggesting the presence of viable cells.
To further confirm the biocompatibility of the hydrogel, after 144 h of incubation, 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT) staining was performed to show mitochondrial activity; samples visualized by optical microscopy confirmed the presence of purple formazan crystals, the index of mitochondrial activity related to cell viability. These data confirmed an excellent compatibility of the hydrogel with NHDF cells, suggesting a good biocompatibility towards healthy skin (Figure 8). To further confirm the biocompatibility of the hydrogel, after 144 h of incubation, 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT) staining was performed to show mitochondrial activity; samples visualized by optical microscopy confirmed the presence of purple formazan crystals, the index of mitochondrial activity related to cell viability. These data confirmed an excellent compatibility of the hydrogel with NHDF cells, suggesting a good biocompatibility towards healthy skin (Figure 8). To further confirm the biocompatibility of the hydrogel, after 144 h of incubation, 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT) staining was performed to show mitochondrial activity; samples visualized by optical microscopy confirmed the presence of purple formazan crystals, the index of mitochondrial activity related to cell viability. These data confirmed an excellent compatibility of the hydrogel with NHDF cells, suggesting a good biocompatibility towards healthy skin (Figure 8). To further confirm the biocompatibility of the hydrogel, after 144 h of incubation, 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT) staining was performed to show mitochondrial activity; samples visualized by optical microscopy confirmed the presence of purple formazan crystals, the index of mitochondrial activity related to cell viability. These data confirmed an excellent compatibility of the hydrogel with NHDF cells, suggesting a good biocompatibility towards healthy skin (Figure 8). The ratio between live and dead cells was investigated through staining with the LIVE/DEAD™ Viability/Cytotoxicity Kit (Thermo Fisher Scientific, Waltham, MA, USA). Figure 9 shows cells seeded on hydrogel and analyzed after 144 h of incubation with a fluorescence confocal microscope after staining. Green fluorescence indicated the presence of viable cells with a good morphology, demonstrating the excellent biocompatibility of the hydrogel.
Pharmaceutics 2021, 13, x 10 of 14 The ratio between live and dead cells was investigated through staining with the LIVE/DEAD™ Viability/Cytotoxicity Kit (Thermo Fisher Scientific, Waltham, MA, USA). Figure 9 shows cells seeded on hydrogel and analyzed after 144 h of incubation with a fluorescence confocal microscope after staining. Green fluorescence indicated the presence of viable cells with a good morphology, demonstrating the excellent biocompatibility of the hydrogel.
Discussion
L-lysine is not reactive enough to directly crosslink HA and a coupling strategy must be designed. We decided to employ the well-known N-(3-Dimethylaminopropyl)-N'ethylcarbodiimide hydrochloride (EDC) and N-hydroxysuccinimmide (NHS) coupling approach, that has already been used for several years, for the crosslinking of amino acids. The mechanism of reaction is reported in Figure 10 and involves the activation of the carboxylic group by the EDC with the formation of an O-acylisourea ester, the formation of an NHS ester with the elimination of an isourea byproduct and, finally, coupling with the amino group that leads to the formation of an amide bond. The byproducts of the reaction are non-toxic and easy to remove through dialysis, while the amide bond is extremely stable in physiological conditions.
Discussion
L-lysine is not reactive enough to directly crosslink HA and a coupling strategy must be designed. We decided to employ the well-known N-(3-Dimethylaminopropyl)-Nethylcarbodiimide hydrochloride (EDC) and N-hydroxysuccinimmide (NHS) coupling approach, that has already been used for several years, for the crosslinking of amino acids. The mechanism of reaction is reported in Figure 10 and involves the activation of the carboxylic group by the EDC with the formation of an O-acylisourea ester, the formation of an NHS ester with the elimination of an isourea byproduct and, finally, coupling with the amino group that leads to the formation of an amide bond. The ratio between live and dead cells was investigated through staining with the LIVE/DEAD™ Viability/Cytotoxicity Kit (Thermo Fisher Scientific, Waltham, MA, USA). Figure 9 shows cells seeded on hydrogel and analyzed after 144 h of incubation with a fluorescence confocal microscope after staining. Green fluorescence indicated the presence of viable cells with a good morphology, demonstrating the excellent biocompatibility of the hydrogel.
Discussion
L-lysine is not reactive enough to directly crosslink HA and a coupling strategy must be designed. We decided to employ the well-known N-(3-Dimethylaminopropyl)-N'ethylcarbodiimide hydrochloride (EDC) and N-hydroxysuccinimmide (NHS) coupling approach, that has already been used for several years, for the crosslinking of amino acids. The mechanism of reaction is reported in Figure 10 and involves the activation of the carboxylic group by the EDC with the formation of an O-acylisourea ester, the formation of an NHS ester with the elimination of an isourea byproduct and, finally, coupling with the amino group that leads to the formation of an amide bond. The byproducts of the reaction are non-toxic and easy to remove through dialysis, while the amide bond is extremely stable in physiological conditions. The byproducts of the reaction are non-toxic and easy to remove through dialysis, while the amide bond is extremely stable in physiological conditions.
Although similar approaches have already been attempted in the past by other authors [29][30][31], they were not successful in providing a simple, flexible and straightforward method. On the contrary, the studies often involved the use of lysine methyl ester during synthesis, followed by a second step of hydrolysis, which leads to the release of highly toxic methanol. This is necessary because, in the reported conditions, the carboxylic group in the alpha carbon of the L-Lysine would be activated and react to form an ester bond with HA, which is not stable in physiological conditions. Other methods also involve performing the reaction at higher pH and temperature, where HA is less stable, long purification processes that employ organic solvents or even destroying the 3D network to allow an easier handling of the gel. Although sterility is mandatory for the medical use of a hydrogel, the behavior of an HA hydrogel stabilized with L-Lysine after autoclaving has not been reported in the literature.
Considering that steam sterilization is a stressful process for HA-based materials, even when crosslinked, the physico-chemical characterization and biological evaluation of bionic HA hydrogel stabilized with L-Lysine formulations have been performed for autoclaved samples.
On the contrary, the synthetic strategy that we report is a simple, one pot synthesis and requires only one purification step, dialysis against PBS at physiological pH, and we show that the hydrogel is stable even after autoclaving.
To optimize the crosslinking degree of the final product, several parameters must be tuned, such as temperature, order of addition of the reagents, concentration, reaction time and pH during the hydrogel preparation. The last parameter is of the utmost importance, not only because HA degradation is very sensitive to pH, but also because the EDC/NHS activation of the −COOH group works better at a pH between 5 and 6, while a basic pH greatly reduces the yield of the reaction by accelerating the hydrolysis of the intermediate. L-lysine coupling efficiency also depends on pH but, on the contrary, it is most favorable at a slightly basic pH, where the amino groups are unprotonated. As shown by the optical microscopy and SEM pictures reported in Figure 3, the hydrogel is a complex 3D microscopical structure made of interconnected pores that allows the diffusion of oxygen and nutrients inside the hydrogel and the proliferation of the cells.
Rheological studies were performed to demonstrate the efficiency of the crosslinking process, after autoclaving, and to characterize the mechanical behavior of the hydrogel under different conditions. A complete rheological characterization is crucial not only because it is necessary to understand the behavior of the hydrogel under stress during injection or after implantation, but the mechanical properties of the gel also influence cell growth and proliferation, especially for applications in tissue engineering, bone repair and osteo-articular fields [4,[32][33][34].
The standard oscillatory protocols show that the crosslinking process is effective and all the samples have a loss module compatible with a gel-like material (i.e., G > G and tan δ < 1). The absolute value of G increases with the concentration of HA, as expected, and is lower for the less crosslinked HA20L sample in comparison to HA20.
The frequency sweep experiment ( Figure 1) further demonstrates that the material has a gel-like behavior at all frequencies, and no crossover points are observed.
To further understand the behavior of the hydrogel in more diverse conditions, we also performed compression and traction experiments. After several cycles of compressions (Figure 2), that mimicked what can happen when the hydrogel undergoes deformations by muscular activity or other physiological movements, the hydrogel still maintained its structural integrity even at a high frequency. The elastic modulus of the new hydrogels, reported in Table 2, are generally lower compared to a similar product in the market, making the hydrogel easier to stretch with a lower force. This is especially important if the hydrogel has to be injected into the dermis or a joint, because it means that it can better adapt to the surrounding tissues and comply with the natural stress produced by movement without causing any mechanical strain that may be uncomfortable for the patient.
New HA-based formulations for medical applications require preclinical studies to evaluate their safety and stability, and to guarantee patient compliance for the following study phases.
In our study, the biocompatibility of the lysine-crosslinked 2.5% HA hydrogel (HA25) was evaluated over a period of 144 h, after direct contact of the material with a normal human dermal fibroblast (NHDF) cell line. The morphological observation of the cells in direct contact with the hydrogel confirmed the characteristic elongated and healthy shape at each timepoint investigated (0-24-48-72-144 h). In order to collect further evidence of the NHDF cell line's healthy state, MTT staining was performed (cell viability indicator) until 144 h of incubation. In this experiment, the yellow tetrazolium salt 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide was reduced to the purple dye formazan by the metabolic activity of the living cells, in particular by NAD(P)H-dependent oxidoreductase enzymes located in the mitochondria.
At all the analyzed timepoints, the formazan crystals observed confirmed a good mitochondrial activity of the cells as an index of cell viability.
To further confirm the biocompatibility of the material, the ratio between live and dead cells was investigated through staining with a LIVE/DEAD™ Kit and confocal microscopic analysis. This experiment was based on the use of a mixture of two dyes, calcein-AM and ethidium homodimer-1, to evaluate the cells' membrane integrity and the activity of esterases, a ubiquitous class of intracellular enzymes. The acetomethoxy derivative of calcein (calcein-AM) is a non-fluorescent dye that, upon hydrolysis of the acetomethoxy group induced by the activity of esterases, releases the green fluorescent dye calcein, with excitation and emission wavelengths of 495/515 nm. The second enzyme, ethidium homodimer-1, interacts with nucleic acids to form a red-emitting complex (excitation at 527 nm, emission at 624 nm). However, it can cross the healthy cellular membrane because of its positive charge, and thus can be used as an indicator of a damaged membrane. The confocal imaging after the LIVE/DEAD™ staining showed an imbalance in the signal in favor of green fluorescence (Figure 9), confirming again that the cells were in good condition and demonstrated the excellent biocompatibility of the hydrogel, and its safety and suitability for in vivo applications.
The degradation kinetic of the hydrogel is another crucial parameter that must be evaluated, in order to understand its lifetime in the body after implantation (for example, in aesthetic medicine or tissue regeneration applications) but also to evaluate its reversibility upon minimally invasive injection of hyaluronidases. Even if one of the purposes of the crosslinking process is to increase the HA hydrogel's resistance to endogenous hyaluronidases, the physicochemical features of the polymer must allow rapid degradation if inappropriate applications occur [35]. Therefore, understanding the susceptibility of this new material to hyaluronidase-mediated degradation is a valid way to complete its safety assessment and its chemical characterization.
The sensitivity to bovine hyaluronidase type I-S was investigated by an in vitro assay under highly controlled conditions, as described in the previous sections.
The results obtained show that the HA20 and HA30 hydrogels reach 82% and 70% maximum degradation, respectively, if placed in contact with 6080 U/mL of hyaluronidase enzyme for a period of 168 h (Figure 4).
The HA concentration does not interfere with the final percentage of degradation, but it is related to the initial hydrogel degradation, while T 50% (time to reach 50% of the maximum degradation value) is higher for HA30 than HA20. The amount of N-acetylglucosamine (NAG) released at different contact timepoints with the hyaluronidase is statistically significant at each timepoint, except for the early ones (1/3/6 h) and the later ones (120-168 h), when a plateau is probably reached.
HA30 does not show a significant increase in NAG release from 48-72 h. Altogether, the results obtained demonstrate a good susceptibility of the lysine-crosslinked HA hydrogel in respect to the exogenous hyaluronidases.
Conclusions
To meet the needs of a global aging population for quality of life, regenerative medicine is seeking biomaterials with a good safety profile for supporting/boosting the regeneration of impaired tissues or function, triggered by degeneration or injury.
We developed a technology platform that can be scaled up to an industrial level, to synthesize a lysine-stabilized bionic hydrogel based on HA for regenerative medicine.
We are able to manufacture hydrogels with customized composition and mechanical properties for the intended use, which can be applied as an injectable, film, topical or scaffold material. The bionic hydrogel mimics biological matrixes, offering an excellent substrate for cell growth and proliferation, with tunable resorption time. The applications in regenerative medicine range from dermal restoration of facial aesthetics, joint health, dentistry and women's health to wound healing and tissue engineering. The preclinical tests are ongoing with excellent results to date. Clinical evaluation of safety and efficacy aims registration as class III medical devices for wound healing and other medical applications.
Patents
De Oliveira Barbosa Nachbaur, L. Alonci G. Funding: Qventis GmbH acknowledges the financial support provided by the Economic Development Agency Brandenburg (WFBB) with the grants GI85010317 and AI85034638 (Funding innovative and innovation assistant) from European Regional Development Funding (ERDF) and European Social Funding (ESF).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable. Data Availability Statement: All the data are available from the authors upon request. | 8,597.4 | 2021-08-01T00:00:00.000 | [
"Biology",
"Materials Science",
"Engineering"
] |
Autoregressive Integrated Moving Average vs. Artificial Neural Network in Predicting COVID-19 Cases in Malaysia
On March 11
INTRODUCTION
An outbreak caused by a new coronavirus was declared by the Chinese government in December 2019 (Saba & Elsheikh, 2020). The virus was first detected from China and has since spread across the globe. On March 11, 2020, the World Health Organisation (WHO) officially announced the outbreak as a pandemic (Katris, 2021). The novel corona virus , a variant of SARS and MERS, began its journey in the Wuhan Province of China on January 21, 2020, and since then has spread to almost every country in the world. As the first cases reported in mid-January in Japan, South Korea, and Thailand, the Government in Chinese agreed to take measures by announcing an emergency lockdown starting on 23 rd January, with compulsory self-isolation and prohibiting travels out of the region (Sweeny et al., 2020). As reported by the World Health Organisation (WHO) on 10 th April 2020, confirmed cases and fatalities increased significantly and exceeded more than 1 million positive cases and more than 100,000 deaths within one month from the declaration date (Saba & Elsheikh, 2020).
On 25 th January 2020, Malaysia declared its first case of Covid-19 after testing close contact with positive cases of Chinese nationals arriving from Singapore in Malaysia (Muhamad, Zainon, Nawi & Ghazali, 2020). There were 190 new confirmed cases of Covid-19 in Malaysia as of March 15, 2020, taking the overall total of positive cases to 428, rendering it the most infected in Southeast Asia. This was the highest in the initial period of the pandemic in Malaysia (Kamaludin et al., 2020). The early move to reduce the positive cases initiated by Malaysia is Movement Control Order (MCO) (Muhamad et al., 2020). MCO's measures included a full ban on people leaving their homes or attending mass gatherings, as well as a restriction on all domestic and international travel. Academic institutions, as well as public and private buildings, were all shut down. The Royal Malaysian Police was called in to assist with the enforcement of the restrictions during this phase.
Despite of the continuous implementation of movement control order, the number of affected cases is not decreasing. Many concerns are looming over the spread of Covid-19 with the number of positive cases keep reaching higher per day. The number of people who will be infected in the upcoming days is keep being questioned every day. Is the curve will keep rising or gets flattened? Are there any mathematical models that could give a solution? Under the circumstances, it is very important to predict potential trends of this diseases so that the government, public health, as well as all citizens could be better prepared to deal with an upcoming emergency.
METHODOLOGY
In this study, two forecasting techniques are utilized to determine the most effective model for predicting upcoming Covid-19 cases: The first is Autoregressive Integrated Moving Average (ARIMA) modelling, and the second is Multilayer Perceptron Neural Networks (MPNN) modelling using an artificial neural network (ANN). From March 1, 2020, to March 29, 2021, 394 observations of daily Covid-19 instances in Malaysia were gathered from an online database www.kaggle.com.my.
Autoregressive Integrated Moving Average (ARIMA)
Autoregressive integrated moving average (ARIMA) is a model derived from the Box-Jenkins methodology. It was initially found in 1976 by Gwilym M. Jenkins and George E. P. Box (Lazim, 2011). ARIMA model have been used in various fields such as economics, social science, epidemiology, medicine, etc. Fattah, Ezzine, Aman, Moussami and Lachhab (2018) has proposed a study on predicting the demand of food company by using ARIMA model and Apergis, Mervar, and Payne (2017) was conducting research to determine the most appropriate model for creating precise predictions of tourists' arrival in Croatia.
In a study conducted by Alzahrani, Aljamaan, and Al-Fakih (2020), four time-series models are applied. The main purpose of the study is to observe and forecast the spread of Covid-19 in Saudi Arabia by using historical data of daily cases. The Box-Jenkins Methodology was used to conduct the study: Autoregressive (AR), Moving Average (MA), Autoregressive Moving Average (ARMA), and ARIMA models, which are synonymous to find the best-fit model. The study concluded that ARIMA is the most suitable model to be applied for prediction purposes. The following are three major stages involved in determining a suitable ARIMA model: 1) Model identification, 2) Model estimation and diagnostic checking, and 3) Model application.
Step 1: Model Identification The first stage in the development of an ARIMA model is the process of determining three parameters 'p', 'd', and 'q' in the model for an appropriate form of ARIMA(p,d,q). The parameter 'p' is determined by an Autoregressive (AR) process. It refers to the number of lagged terms of the dependent variable. Next, parameter 'd' denotes the number of differencing orders which require transformation from non-stationary series to a stationary time series. Lastly, parameter 'q', which is determined by Moving average (MA), refers to the number of lagged time or the order of moving average.
Prior to identifying a model, we must decide whether the series is stationary. If the stationary condition is not fulfilled by the series of data, a process of differentiation is necessary to transform it to a stationary series. There are four procedures to look at to check stationarity. Firstly, a time series data is plotted to see whether the series is constant around the mean value. Secondly, the correlograms, namely autocorrelation function (ACF) and the partial autocorrelation function (PACF), are analyzed. Finally, the two most popular approaches to test the unit root hypothesis, namely Augmented Dickey Fuller (ADF) and Kwiatkowski-Phillips-Schmidt-Shin (KPSS), are computed. Below is the ADF test by using the Ordinary Least Square (OLS) procedure to estimate the model, where is the number of lags for ∆ which ∆ = − −1 , is white noise with mean zero and variance 2 . The null hypothesis for this test, which presupposes that the series is not stationary, must be disproved. On the contrary, the null hypothesis for the KPSS test assumes that the data are stationary. Therefore, we wish to avoid rejecting the null hypothesis for this test. If both tests fail to get the series stationary, the differencing process is required until the data become stationary. The number of differencing processes is represented by the parameter 'd' in the ARIMA model. After successive unit root tests, we are able to identify all of the 'p', 'd' and 'q' parameters in ARIMA models.
Step 2: Model Validation and Diagnostic Checking
Model creation is a collaborative and iterative process. Numerous alternative models are evaluated before choosing the one that generally performs "best" using a criterion like Akaike's Information Criteria (AIC) (Chatfield & Xing, 2019). In this study, the fitness of an ARIMA model is evaluated using AIC. The ARIMA model is most appropriate when the AIC is low. Mathematically, it is formulated as: is the total of AR value plus MA value. is the number of observations in the data.
2 is a function to avoid overfitting the model. Besides AIC, another common statistical measure used to validate the ARIMA models is the Ljung-Box statistic which is given as, is the number of observations in the time series data. ℎ is the maximum lags being tested. is the order of AR terms. is the order of MA terms. is the ℎ sample autocorrelation of the residual terms. is the degree of differencing.
Step 3: Model Application
Once the model's fitness has been confirmed, it is then ready to be utilized to produce prediction values, where the accuracy between the actual output and the anticipated output is compared. This final stage can be done by using RStudio. The general steps necessary in creating ARIMA modelling are shown in Figure 1.
Multilayer Perceptron Neural Network (MPNN)
The second approach employed in this study is known as a Multilayer Perceptron Neural Network (MPNN), which is a feed-forward neural network type that is developed from the simple perceptron. It may represent non-linear functions by incorporating one or more hidden layers. A wide range of learning algorithms have been proposed to train multilayer perceptron networks. The back-propagation algorithm was the initial learning method designed for this purpose, and it is now used in nearly all business applications (Rodrigues & Carpinetti, 2019). Ranjan, Majhi, Kalli and Managi (2021) have used Multilayer Perceptron model to predict gross domestic product (GDP) in eight countries. Their outcome shown that the MPNN model was able to accurately forecast GDP figures with a lower mean absolute percentage error (MAPE) as well as performing well for solving economic-related issues. Additionally, Slimani, Sbiti and Amghar (2019) have compared a variety of neural network models such as Perceptron, Adaline, Radial Basis Function (RBF), NoProp, and Multilayer Perceptron (MPNN) to solve the traffic jam problem. With the smallest error in both training set and test set, Multilayer Perceptron (MPNN) generated the best accuracy forecast compared to other models. Another study has done by Zealand, Burn, and Simonovic (2000) for streamflow forecasting. The study has proved that the neural network model gave the best accuracy result compared to The Winnipeg Flow Forecasting System (WIFFS) model with the lowest root mean squared error (RMSE). They concluded that data-driven methods like neural network are suitable for handling complex problems.
In this work, the Alyuda NeuroIntelligence software is used to model multilayer perceptrons. Prior to obtaining the best neural network model, six main steps need to be executed. The stages are depicted in Figure 2. Step 1: Analyzing data The data must be checked before being imported into the Alyuda programme to remove data anomalies such as outliers or drifts that would adversely affect the network's performance. The dataset will be examined and divided into three parts, namely the training set, the testing set, and the validation set. To reduce overfitting, more data is used for the training set in the ratio of 75:25.
Step 2: Pre-processing data The outcomes of pre-processing data will be checked in stage two. The data needs to be clean before entering the network. Data cleaning involves the process of fixing or removing incorrect, corrupted, incorrectly formatted, duplicate, or incomplete data within a dataset.
Step 3: Designing network architecture Designing the network requires the decision of the number of nodes in the hidden layer and the number of architectures. By using Kolmogorov's Superposition Theorem, the number of nodes can be calculated mathematically as: represents the total of nodes in the hidden layer. While indicates the total of nodes in the input layer.
Analyzing data Pre-processing data Design network architecture Training the network Testing the network Validation Step 4: Train network At this stage, the optimal algorithm to employ for network training will be determined based on the model that gives the lowest value of absolute error. There are six different algorithms available in Alyuda NeuroIntelligence: Conjugate Gradient Descent, Quick Propagation Quasi-Newton, Levenberg-Marquardt, Limited Memory Quasi-Newton, Batch Back Propagation, and Online Back Propagation. When choosing these training algorithms, a lot of aspects are considered, including interpretability, the quantity of data points and features, the data format, and many more.
Step 5: Test network At this stage, the data will be validated and tested. Once the training phase is completed and the model parameters have been modified, the validation set is used to assess how well the model fits. Overfit and underfit scenario are identified by checking validation metrics such as accuracy and loss. The model is underfit when the validation loss is decreasing, and the model is overfit when validation loss is increasing. This study uses dropout layers to handle overfitting situation. In the testing phase, the targeted and predicted values are evaluated to assess the model's performance and identify the error that exists between them. As a result, the dataset's mean absolute error (MAE) is obtained by using the equation below.
is prediction, is target value and is total number of data points.
FINDINGS AND DISCUSSIONS Autoregressive Integrated Moving Average (ARIMA) model
The data is split into two sections during the model-building process: estimation and evaluation. The estimation component, or training data is used to fit the model, while the evaluation component, or test data, is used to evaluate the model's accuracy. With 295 training data and 99 testing data, the data splitting percentage is 75 percent versus 25 percent. All the training data (1 st March 2020 until 20 th December 2020) are transformed into a time series data type. Figure 4 shows the time series plot of Covid-19 cases in Malaysia. The series shows an upward trend over the period from March 1, 2020, until December 20, 2020, indicating the data series is not stationary. To satisfy the stationarity requirement, the trend must be eliminated. Thus, the process of differencing is required until the data series is stationary.
Figure 4: Time Series Plot of Covid-19 cases in Malaysia
First order differencing of the time series data is implemented, and the results obtained are shown in Figure 5, Figure 6, and Figure 7. From the result of Figure 4, it can be seen that the series in the first difference fluctuates randomly around zero value. Together, the present findings confirm it does not show growth or decline over time. More specifically, there is no presence of trend component. Therefore, the series can be said stationary. To verify its stationarity, the two statistical tests ADF and KPSS are conducted. The results of both tests after the first order differencing are presented in Table 1. Both tests show that the data series is stationary. Therefore, further differencing the time series is no longer required and we adopt d=1 for the ARIMA(p,d,q) model since the process of differencing is only performed once.
The sample autocorrelation function (ACF) and the sample partial autocorrelation function (PACF) are the main tools to identify the initial ARIMA model for a given stationary time series. The parameters "p" and "q" in ARIMA (p,d,q) are guessed by the number of significant spikes that exceed the two significant limits (blue lines) in autocorrelation function (ACF) and partial autocorrelation function (PACF) plots in Figure 6 and Figure 7 respectively. According to Figure 6, there are significant spikes at lag 1, lag 2, lag 11, lag 12, lag 15, lag 17, lag 18, lag 19, and lag 21. This indicates moving average (MA) term of order is 9 (q=9). Referring to Figure 7, the significant spikes that extend beyond the limits are at lag 1, lag 2, lag 3, lag 4, lag 6, lag 10, lag 14, lag 15, lag 16, lag 18, lag 21, and lag 24. This indicates an autoregressive (AR) term of order 12 (p=12). As a result, the finding suggests ARIMA (12,1,9) as an initial model. But ARIMA (12,1,9) might not be the best fit. Therefore, in order to find the best model, other possible models are identified by considering every possible combination 'p' and 'q'.
A well fitted model produces the lowest Akaike's Information Criteria (AIC) and the residuals obtained are expected to be independently distributed. There are 25 models based on the Ljung-Box test, with independently distributed residuals (the errors are white noise) as shown in Table 2. However, a comparison of their AIC values is performed in order to select only one best model out of these 26 well-specified models. Based on the smallest value of AIC, the result points towards ARIMA (4,1,5). Thus, ARIMA (4,1,5) is the best fit model among the others. The mean absolute error of ARIMA (4,1,5) is then calculated to assess the accuracy of its forecasts, and the result is 1096.799.
Multilayer Perceptron Neural Network (MPNN)
The dataset for this study is partitioned into three parts: training set, testing set, and validation set as shown in Table 3. Since the neural networks can only process numeric inputs, all the data are in numerical format. Neural networks also require the input to be scaled in a consistent way. Therefore, the data is standardized into a distribution with a mean of 0 and a standard deviation of 1, through normalization process. The structure of neural networks consists of input, hidden neurons, and output. The most effective network in this study is [2-1-1] based on the minimum AIC. [2-1-1] refers to a network design with two inputs, one hidden node, and one output chosen from the 10 possible network architectures as shown in Table 4. Next, the network architecture [2-1-1] is trained with seven training algorithms as shown in Table 5. Based on the absolute error of these training algorithms, Quasi-Newton shows the smallest training absolute error. Hence, this algorithm is chosen to be applied on the [2-1-1] network in the testing phase. In the final phase (testing phase), the performance of the [2-1-1] neural network model is evaluated by using the testing dataset. Based on the testing result in Table 6, we may conclude that Multilayer Perceptron Neural Networks are appropriate for time series forecasting purposes, given the low mean absolute error of 334.59. Comparison between ARIMA and MPNN model Figure 8, Figure 9, and Table 7 provide graphical and statistical results for evaluating the forecast performance of ARIMA and MPNN. Based on the Table 9, the mean absolute errors (MAE) for the MPNN model are the lowest, demonstrating its better performance over the ARIMA (4,1,5) model. Additionally, the outcomes of Figures 5 and 6 also support this conclusion.
Forecasting the future Covid-19 cases in Malaysia
Next, the best MPNN model [2-1-1] is used to predict how many new cases there will be in the next 30 days. Figure 9 shows an upward trend in the number of instances. The number of positive cases is steadily increasing beginning in April of 2021. There are several potential causes for this increase in the number of cases, but the establishment of community clusters is most certainly one of them. Several ideas that can be utilized to enhance future research. Firstly, the same techniques should be used for real-time data updates. In terms of forecasting accuracy, a larger number of data would provide more accurate model forecasting. Finally, as suggested by Aggarwal et al. (2020) and Phan and Nguyen (2020), hybridization of ARIMA-ANN should be good to be proposed. | 4,239.2 | 2022-09-30T00:00:00.000 | [
"Computer Science",
"Medicine"
] |
Risks of long-term leasing transactions for construction industry development
The article examines potential long-term leasing risks for development of the construction industry, characterizes leasing as a form of investment. Leasing objects of the construction industry were identified. Methods for minimizing risks were considered. Leasing companies and large construction companies rarely lease out equipment to private, small-sized construction companies facing general and specific risks. The degree of risk depends on the leasing object (e.g., equipment that cannot be disassembled, rare equipment, expensive equipment which is difficult to sell at an affordable price). The problem to be solved is lack of owned funds for carrying out innovation activities. The article describes funding methods for innovation activities of manufacturing enterprises. It has been found that the most common funding method is leasing. Innovation activities of enterprises are a number of measures aimed to search for, commercialize and implement scientific knowledge, new technology and inventions. Currently, innovation activities influence the Russian economy by enhancing efficiency of resource management, creating new industries, infrastructure and jobs. Under the global competition, the effects of innovation activities on national development and competitiveness are becoming more intensive.
Introduction
Leasing is a rent and a financial service. Under the leasing agreement, property is acquired for the needs of a specific tenant.
Leasing transactions have gained the greatest popularity among businesses. They arre used in various areas, including the construction industry. In fact, leasing allows a legal or natural person to use equipment without purchasing it. Leasing objects in the construction industry are as follows: special equipment, narrow-profile equipment, computer equipment and software.
According to experts, about 50% of fixed assets of the construction industry is worn out which requires their renewal.
The objects of leasing operations in the construction sector are as follows: industrial equipment, cars, cranes, containers, vessels, etc. For construction companies lacking equipment and machines, leasing operations can be an effective mechanism for upgrading the machinery and equipment. Small enterprises and small contracting construction organizations need various types of equipment for several monthsto construct an object. They cannot afford their own repair base, diagnostics and maintenances specialists, etc.
The relevance of this study is due to the fact that leasing transactions generate a large number of risks that can affect the financial development of the lessor and the lessee. The main way to minimize risks is to predict them.
Materials and Methods
The article describes characteristics of leasing, identifies potential risks of rental transactions in the construction industry. The way to solve this problem is to predict these risks. The research methods used in the study are analysis and synthesis.
In the construction industry, leasing is used not only for acquisition of vehicles and equipment. It is used when the lessee wants to receive an enterprise which will be further transferred to his/her ownership.
There are different forms of financing of innovation activities: -leasing involves purchasing of capital assets by enterprises or expensive goods by individuals. It is a long-term rent which contains an option to purchase the leased property at a bargain price and decrease a tax load (Ries, 2011) -factoring is a financial transaction in which a bank purchases debt recovery rights.
-forfeiting is a financial transaction involving the purchase of receivables from exporters by a forfeiter.
Despite the advantages of leasing transactions as an important financing source, there are some obstacles to leasing development in Russia: -short loan terms and high loan rates; -high tax rates; -if a company purchases equipment at full cost, it lacks sufficient seed capital; -lack of information about offers of leasing services; -underdeveloped leasing market infrastructure; -financial risks of leasing activities; -macro-economic and political uncertainty in Russia. Currently, leasing is a promising and flexible economic tool attracting investment funds in real economic sectors, providing support for development of national industries, ensuring reliable longterm income of credit institutes, and providing business support.
The most attractive industries for Russian leasing development are transport (e.g., air transportation), machine building, agriculture, and small business.
To eliminate obstacles to leasing development, the following measures can be taken: -Market restructuration and reforming to control risks and prevent defaults in the leasing industry.
-Expansion of benefits for long-term financing of leasing deals. -Insurance.
-Attraction of foreign investment in the leasing industry.
-Development of the leasing market infrastructure (e.g., leasing counselling, qualified staff in the leasing industry).
Currently, new leasing companies appear. They have high positions in the national market. The current level of innovation development in Russia is lower than the one in developed countries. Decreasing volumes of public funding, the lack of owned funds development strategies cannot be compensated for private investment [4]
Results
Leasing companies are ready to solve the problem on a turnkey basis, i.e., to provide workshops, factories or other enterprises with all the required equipment.
This transaction helps the lessee avoid the cost of delivery, customs clearance and equipment installation. The leasing company may acquire facilities at the construction stage under the purchase and sale contract for further leasing of these facilities. At the same time, the leasing company cannot put an unfinished facility into service because it is not its owner.
Construction companies often conclude leasing agreements. The advantages of leasing transactions for the construction industry are as follows: Reduced lump-sum expenses. Advance payment under the leasing agreement ranges from 10 to 20%, and in some cases the acquisition of property can be fully financed by the leasing company.
Tax benefits. Due to the accounting for lease payment, the company can reduce the income tax. The VAT is reimbursed.
Flexible payment schemes. The payment schedule can take into account wishes of the client and specifics of the business. For example, it may take into account the seasonal nature of the lessee's profit.
Accelerated depreciation. It is possible to reduce tax payment and update outdated equipment and machines.
Along with these advantages, leasing transactions have disadvantages in the financial and credit spheres and cause accounting problems: 1. The leasing agreement is complex due to a larger number of participants.
2. The disadvantages for the lessee are as follows: -in case of the financial leasing, lease payments do not stop until the end of the contract, even if scientific and technical progress makes the leased property obsolete; -the lessee does not benefit from an increase in the residual value of the equipment; -the tax-based return international leasing is often loss-making for the country of the lessor; -in case of international multicurrency leasing transactions, there are no full guarantees against currency risks, i.e. problems are transferred from one participant to another.
In addition to the advantages, there are leasing risks. One of the main problems is the need for investment (financial risks).
Force majeure is another risk that can generate significant problems. The leasing risks are presented in Table 1. Marketing risks Impossibility to transfer the equipment to the tenant 2 Accelerated obsolescence risks Since equipment is subject to physical and moral obsolescence, scientific and technical development has to be taken into account 3 Price risks Constantly changing market conditions have a direct impact on the price of the leased asset and potential loss of profit 4 Decommissioning risks This risk can be avoided by performing insuring leased property.
Unbalanced liquidity risks
Financial losses due to the lessor's inability to perform obligations 6 Default risks Inability and impossibility of the lessee to perform obligations under the lease agreement 7 Interest risks Profit deficiency due to changes in interest rates 8 Political risks Unregulated and unpredictable public sector policy which affects all activities, including leasing.
Despite the risks and shortcomings of leasing transactions, leasing is developing and gaining momentum. According to RAEX ("Expert RA"), in 2017, the volume of leasing transactions in the construction and road-building equipment sector increased by 49% and amounted to 76 billion rubles. In 2018, the demand for construction equipment was growing. In the Russian construction and road- An increase in transactions was due to the preferential special equipment leasing program developed by the Ministry of Industry and Trade. A discount of up to 10% of the cost of domestically produced equipment is offered by reducing the advance payment.
The government implemented most of the strategic initiatives for industrial development, renewal of production assets (including construction equipment) through leasing companies and subsidized programs for acquiring fixed assets. The leasing companies act as guides for the government support measures, since subsidies are given to the lessee.
Various discounts from equipment dealers and joint programs of leasing companies and manufacturers make the leasing profitable. Support programs developed by large automobile producers make it possible to lease construction equipment with zero appreciation, minimum monthly payment or minimum advance payment. Minimum advance payment is in great demand when leasing special equipment. These offers can significantly reduce the burden on corporate clients leasing equipment. Therefore, they are always popular.
In the first half of 2018, leasing construction equipment was used for road construction. On the eve of the World Cup 2018, road and construction companies concluded a lot of leasing agreements in order to repair city roads.
Construction equipment leasing agreements are often concluded by leasing companies, utilities companies, construction material manufacturers, and agricultural companies.
Excavators, backhoe loaders, tractors are in demand. The lessors usually lease construction equipment. The advance payment is 10% and the contract period is up to 3 years.
According to the forecast, in 2019, the volume of leasing transactions will increase by 25-30%. Innovation enterprises need significant investment funds. Innovation and investment are interrelated in current manufacturing processes regardless of business scales. Investment in innovation activities aims to introduce new technology in company activities. However, it is necessary to understand that project profitability should be a priority. There are two investment purposes: purchasing of innovation products, licenses, patents; development of innovation products. Unfortunately, the investment structure is homogenous enough and depends on its raw material component. In developed countries, innovation activities are funded by public and private organizations. Equipment of the real economy sector has a high degree of wear (more than 60%). Depreciation of fixed assets has always been acute. This problem has gained special significance due to international sanctions against Russia. The import substitution policy of the Russian government faced a number of problems: unwillingness of enterprises to increase production, inability to launch production due to technical problems of industrial enterprises. The Russian is experiencing large-scale re-equipment of production and implementation of new technologies [15] In the construction industry, there are a lot of leasing risks due to the fact that the leasing process is based on the division of labor. Leasing participants are autonomous economic entities with their goals and objectives. This fact makes it difficult to identify risks.
Discussion
To obtain the maximum result, it is necessary for each participant to have a shared goal; otherwise, a "domino effect" may occur: the collapse of one company entails the bankruptcy of companies associated with the bankrupt one.
The leasing risks are interrelated and have direct positive and negative impacts. They can be distributed between participants in forward and opposite directions. The multidirectional effect of leasing risks makes it difficult to identify and assess their influence on individual participants and entire leasing business.
To prevent risky situations with negative effects, it is necessary to predict these negative situations and take measures to increase the efficiency of the leasing sector of the economy.
Risk classification plays a special role in identifying risks of the leasing company. It is based on causal relationships. Risks of one participant directly or indirectly influence other participants at the micro and macro levels. It is necessary to take into account the multivariate nature of risk assessment and consequences of decision-making at various levels. The classification determines the role of each specific risk which contributes to development of management decisions.
When determining leasing risks, it is necessary to use a step-by-step method, since the risks are distributed in time and space. This type of distribution causes negative cumulative effects.
The step-by-step planning of the multi-step process makes it possible to find an optimal solution and manage risks in the leasing sector taking into account time of their occurrence and coordinates.
The solution to this problem is risk management, risk minimization.
Conclusion
The main incentive for the use of leasing in the construction industry is improvement of the efficiency of investment. Investments is fixed assets required for economic activities of construction companies. The lesser has a sales channel for leasing, which means that the lesser can perform leasing activities at a faster rate. Leasing transactions are beneficial for the economy. The leasing mechanism is improving. A special role in identifying risks of a leasing enterprise is assigned to risk classification, which is based on a causal relationship. The risks of one participant generate the risks of another participant who has a direct connection and indirectly the risks of participants who have a direct relationship, and who influence the formation of the result, at the micro and macro levels. It is necessary to take into account the multivariate nature of risk assessment and the consequences of decision-making at various levels. Classification allows you to set the place of each specific risk, which contributes to the development of management decisions.
Risk classification plays a special role in identifying risks of the leasing company. It is based on causal relationships. Risks of one participant directly or indirectly influence other participants at the micro and macro levels. It is necessary to take into account the multivariate nature of risk assessment and consequences of decision-making at various levels. The classification determines the role of each specific risk which contributes to development of management decisions. When determining leasing risks, it is necessary to use a step-by-step method, since the risks are distributed in time and space. This type of distribution causes negative cumulative effects.
Innovation activities are crucial for sustainable development. To this end, it is important to monitor investment in knowledge, technology and ideas contributing to innovation development.
Investment management tools in innovation enterprises should be system-based which requires development of an investment management system. Even under limited investment resources, an efficient investment management system can create possibilities for national innovation sector development.
Under existing economic conditions, mobilization of new development sources and utilization of global innovation opportunities are a priority for all subjects concerned [4].
The step-by-step planning of the multi-step process makes it possible to find an optimal solution and manage risks in the leasing sector taking into account time of their occurrence and coordinates. | 3,371.2 | 2019-11-28T00:00:00.000 | [
"Business",
"Economics"
] |
Interpolation Approach to Hamiltonian-varying Quantum Systems and the Adiabatic Theorem
Quantum control could be implemented by varying the system Hamiltonian. According to adiabatic theorem, a slowly changing Hamiltonian can approximately keep the system at the ground state during the evolution if the initial state is a ground state. In this paper we consider this process as an interpolation between the initial and final Hamiltonians. We use the mean value of a single operator to measure the distance between the final state and the ideal ground state. This measure could be taken as the error of adiabatic approximation. We prove under certain conditions, this error can be precisely estimated for an arbitrarily given interpolating function. This error estimation could be used as guideline to induce adiabatic evolution. According to our calculation, the adiabatic approximation error is not proportional to the average speed of the variation of the system Hamiltonian and the inverse of the energy gaps in many cases. In particular, we apply this analysis to an example on which the applicability of the adiabatic theorem is questionable.
and the inverse of the energy gaps in many cases. In particular, we apply this analysis to an example on which the applicability of the adiabatic theorem is questionable.
Introduction
Adiabatic process is aimed at stabilizing a parameter-varying quantum system at its eigenstate. This process has many applications in the engineering of quantum systems [1,2,3,4,5], and in particular plays the fundamental role in adiabatic quantum computation (AQC) [6,7,8]. The adiabatic theorem [9,10] states that a system will undergo adiabatic evolution given that the system parameter varies slowly.
The validity of the adiabatic theorem has been under intensive studies both theoretically and experimentally since it was proposed, and much of these efforts were devoted to the rigorous description of the sufficient quantitative conditions of adiabatic theorem, and the estimation of the error accumulated over a long time [10,11,12,13]. Once the exact knowledge on the adiabatic process is available, it is straightforward to apply the results to the optimal design of adiabatic control on specific systems [14,15]. The most interesting progress is that the validity of the adiabatic theorem itself has been challenged in the recent decade [16,17,18,19,20,21,22,23,24], both by strict analysis and counter-examples. According to these findings, the errors induced by the adiabatic approximation could accumulate over time despite certain quantitative condition is satisfied [16,17,18,21,22], e.g., when there exists an additional perturbation or driving that is resonant with the system. Particularly as indicated in [21], it is not new that resonant driving can cause population transfer between eigenstates. Also, a proof can be found in [22] stating that only a resonant perturbation whose amplitude gradually decays to zero can result in a violation of a well-known sufficient condition.
In this paper we consider the following process: The initial Hamiltonian is H 1 . The final Hamiltonian H 2 = H 1 + λ∆H is obtained by varying H 1 along a fixed direction ∆H. λ is the parameter which measures the maximum variation of H(t). We define an interpolating function T is the evolution time. We work under the condition that a valid perturbative analysis of the system evolution is available. This often means λ should be smaller than a threshold value. Therefore, our analysis in this paper is not concerned with the adiabatic evolution for a large variation of Hamiltonian. However, our analysis provides a rigorous estimation of the error accumulated during this small-variation evolution for an arbitrarily given interpolation.
Our work is different from the previous works in two ways. First, instead of studying the evolution of the eigenstates and their corresponding probability amplitudes, the mean value of a Hermitian operator is defined as a measure of the error. For example, in the context of adiabatic quantum computation where one wants to prepare the ground state of a target Hamiltonian H 2 ≥ 0, ǫ = H 2 ρt will serve as a good measure of the distance between the real-time state ρ t and the ground state. In this paper we only consider the error accumulated over the entire process, which means we are only interested in ǫ = H 2 ρ T . This operator approach provides another way to look at this time-dependent evolution and sometimes greatly simplifies the calculations. The second difference is that the error can be estimated with a sufficient precision for arbitrarily given interpolating functions. As a result, the parameters which are related to the suppression of the error can be easily identified. For example, we have ǫ = O( λ 2 as λ → 0 in the case of linear interpolation. However for the interpolation in the counterexample [16,22], the scaling of ǫ is not so simple.
This paper is organized as follows. In Section 2, we introduce the model of this paper. In Section 3, we give the estimation of the error for linear interpolation. In section 4, we present the general algorithm to estimate the error for an arbitrarily given interpolating function. We discuss three examples in Section 5. Conclusion is given in section 6.
Definitions and Preliminaries
To begin with, we consider the process starting at t = 0. The system Hamiltonian at t = 0 is H 1 , and the system Hamiltonian at t = T is H 2 = H 1 + λ∆H, λ > 0. ∆H is a fixed operator and so the direction of the variation is fixed. T is the evolution time. We assume H 1 , H 2 , and ∆H are bounded operators throughout this paper.
Let {ω i : i = 1, 2, ...N} be the monotonically increasing sequence of eigenvalues of H 1 , so that ω i ≥ ω j when i > j, and {|i } be the corresponding eigenstates. We denote the energy gap between the ith eigenstate and the ground state as λ i = ω i − ω 1 . Similarly, we define the increasing sequence of eigenvalues of H 2 , {ω For convenience, we also introduce two offset Hamiltonians,Ĥ 1 andĤ 2 . The HamiltonianĤ 1 is defined asĤ 1 = H 1 − ω 1 , i.e., by offsetting the Hamiltonian of the system at t = 0 by a constant operator ω 1 so thatĤ 1 ≥ 0. ByĤ 1 ≥ 0 we meanĤ 1 is positive semidefinite and its the smallest eigenvalue ofĤ 1 is zero. Similarly, we definê H 2 = H 2 − ω ′ 1 ≥ 0 by offsetting the system Hamiltonian by a constant operator ω ′ 1 . Let ρ t denote the system state at time t and let ρ g be the initial state of the system at t = 0. We always assume that ρ g is the ground state ofĤ 1 , and so we have Ĥ 1 ρg = 0.
The transition of the system from H 1 to H 2 can be described using an interpolating function f (t) so that with f (0) = 0 and f (T ) = 1. In particular, the linear change from H 1 to H 2 along the direction ∆H, can be described using the increasing interpolating function f (t) = t T . The measure of adiabaticity is proposed as follows Interpolation Approach to Hamiltonian-varying Quantum Systems and the Adiabatic Theorem4 Definition 1 The distance between the final state and the ground state of H 2 is measured by Obviously, if the evolution is adiabatic, i.e., ρ T is the ground state of H 2 , then we have ǫ = 0. In particular, ǫ is closely related to the fidelity of the final state and ground state in the Schrödinger picture (See Appendix C). A small error ǫ implies a large fidelity.
In this paper we also call ǫ the adiabatic approximation error, as ǫ reflects how well we can approximate the evolution as a perfect adiabatic process.
In this paper we only consider λ such that ρ t , t ∈ [0, T ] can be expanded using Magnus series in the interaction picture. For more details about the expansion in the interaction picture, please refer to Appendix A. If the series expansion is valid in the interaction picture, we can transform back to the Schrödinger picture and write the evolution of the state as (see Appendix A) where we have ||R(t)|| = O(λ 2 ). A sufficient condition for the Magnus series to converge is given by (see Appendix A) Our aim is to estimate an asymptotic behaviour of ǫ provided λ → 0. Furthermore, we will use the obtained estimate to analyze several cases of the adiabatic theorem including those where some difficulties with adiabatic approximation have been encountered.
Adiabatic approximation under linear interpolation of the Hamiltonian
The Heisenberg evolution of the expectation of an observable is written as where H is the system Hamiltonian. Recall that ρ g = |1 1|. Since H 1 |1 = ω 1 |1 , X(t) ρg is a constant of motion under the action of H 1 : for any Hermitian operator X(t).
We will need to study the dynamics of Ĥ 2 ρt = Ĥ 2 (t) ρg in order to solve for ǫ. The time evolution of Ĥ 2 ρt is determined by its generator d dt As we noted before, Ĥ 2 ρ T is exactly zero if ρ T is the ground state ofĤ 2 . If ρ T is not the ground state ofĤ 2 , we can determine the bound on ǫ = Ĥ 2 ρ T from the following equality The error ǫ can be expressed as With the aid of (8), we can exactly calculate ǫ in the case where f (t) defines a linear interpolation, as summarized in the following proposition: Proposition 1 Assume λ 2 > 0 (the ground state of H 1 is non-degenerate) and suppose f (t) = t/T , which corresponds to the linear interpolation of the Hamiltonian. ǫ is of the order O( λ 2 Proof 1 Referring to (11) and (9), we need to compute the difference between (8) and ω ′ 1 − 1|H 2 |1 . First we write (8) as Interpolation Approach to Hamiltonian-varying Quantum Systems and the Adiabatic Theorem6 DenoteH = max f (t)∈(0,1) ||H(t)||. Since is O(λ 3 ), we can further write (13) as Clearly, the term O( λ 2 ) dominates as λ → 0.
Next we will calculate ω The smallest eigenvalue ω ′ 1 of H 2 can be calculated using the first-order time-independent perturbation theory for non-degenerate system.
Assume H 1 is the unperturbed Hamiltonian and the perturbation is λ∆H, then the lowest eigenvalue of the perturbed Hamiltonian H 1 + λ∆H can be written as series in terms of λ and ω 1 [25]: Thus we conclude Comparing (16) and (19), the second order terms cancel and so the error ǫ is estimated by
Error Estimation for Arbitrary Interpolations
The approach derived in the previous section can be easily generalized for arbitrary given continuous interpolating functions. The generalization can simply be done by Interpolation Approach to Hamiltonian-varying Quantum Systems and the Adiabatic Theorem7 replacing the linear interpolation function with the given continuous function f (t) and then recalculating the double integration in (7). The error estimation can easily be obtained from the proof of Proposition 1:
Proposition 2
For an arbitrarily given f (t), the error estimation is given by as λ → 0.
Proof 2 ǫ is still calculated by (11), using H 2 ρ T − H 2 ρg and ω and It must be pointed out that A(T ) is very easy to calculate with the aid of any softwares that can perform symbolic integration, and therefore it is straightforward to apply Proposition 2 to find the error estimation for a given interpolating function, as we are going to do in the next section.
Linear Interpolation: f (t) = t/T
By Proposition 1, the error estimation is ǫ = O( λ 2 T 2 λ 3 2 ) as λ → 0. This error term is determined by λ T which is the average speed of the variation of the system Hamiltonian, and 1 λ 2 which is the inverse of the energy gap between the ground and first-excited eigenstates of H 1 , as λ → 0. In particular, we have Therefore, since the inverse of the energy gap 1 λ 2 is a fixed value, the approximation error ǫ is estimated to be proportional to the square of the average speed of the variation of the Hamiltonian, which is ( λ T ) 2 , as λ → 0.
Replace f (t) with a nonlinear function f (t) = t 2 T 2 in (7) and we recalculate the integral to be By Proposition 2, for sufficiently small λ, the error is estimated to be of order of λ 2 : That is, in contrast to the linear interpolation case, we have This calculation shows that if the evolution speed is infinitely slow, then the system dynamics is adiabatic during t ∈ [0, T ]. However, the scaling of ǫ quad with respect of the square of the average evolution speed λ T is not as simple as in the linear case, where the scaling of ǫ with respect of ( λ T ) 2 is solely determined by the inverse of the energy gaps as λ → 0. In the quadratic case, this scaling is determined by a complex factor
Interpolation with Decaying Resonant Terms
Here we assume a linear interpolating function with an additional oscillating term that gradually decays to zero. That is, where λ c is the oscillating frequency of the perturbation. Ortigoso observed in [22] the inconsistency in the applicability of the adiabatic theorem when the Hamiltonian contains resonant terms whose amplitudes go asymptotically to zero. (7) and we recalculate the integral to be i =1 .
Q 1 is a function of four parameters. Obviously, (29) is singular when λ c = λ i for some i.
To be more precise, we can let λ c = λ i for an arbitrary i and find that the coefficient where Q(T ) is a complicated fraction with T being in its denominator. The error resulting from the ith term is given by as λ → 0. We have The scaling of ǫ i with respect of ( λ T ) 2 is additionally determined by T 2 and T 4 , as compared to the quadratic case. This is where adiabatic evolution may break down even if the average evolution speed is slow. In particular by (32), if one chooses a comparably large value for T in an adiabatic evolution experiment, the adiabatic approximation error may not decrease as expected when one applies a slow evolution speed λ T . In addition, we can compare this case with the quadratic case using the ratio Note that the inner limit is taken for T being constant. Therefore, the rate of convergence considered in this subsection is slower than that in the quadratic or linear case. i.e., ǫ goes to zero as λ → 0 at a much slower rate than in the linear interpolation case or the quadratic interpolation case if T is large. Furthermore, the larger T is, the slower the convergence.
Conclusion
In this paper we provide a rigorous analysis of the time-dependent evolution of Hamiltonian-varying quantum systems. As we calculated, the adiabatic approximation error is not proportional to the average speed of the variation of the system Hamiltonian and the inverse of the energy gaps in many cases. The results in this paper may provide guidelines when applying complicated interpolation for adiabatic evolution.
Appendix C.
The state of the system will remain a pure state during the evolution. Therefore, we can express the final state as ρ T = |ψ ψ| with |ψ = N i=1 c i |i ′ . Using this expression, the error measure ǫ defined by (2) can be written as (C.1) The fidelity of the final state and the ground state |1 ′ is calculated by (C.2) | 3,763.2 | 2015-03-11T00:00:00.000 | [
"Physics"
] |
Plant Peroxisomes: A Factory of Reactive Species
Plant peroxisomes are organelles enclosed by a single membrane whose biochemical composition has the capacity to adapt depending on the plant tissue, developmental stage, as well as internal and external cellular stimuli. Apart from the peroxisomal metabolism of reactive oxygen species (ROS), discovered several decades ago, new molecules with signaling potential, including nitric oxide (NO) and hydrogen sulfide (H2S), have been detected in these organelles in recent years. These molecules generate a family of derived molecules, called reactive nitrogen species (RNS) and reactive sulfur species (RSS), whose peroxisomal metabolism is autoregulated through posttranslational modifications (PTMs) such as S-nitrosation, nitration and persulfidation. The peroxisomal metabolism of these reactive species, which can be weaponized against pathogens, is susceptible to modification in response to external stimuli. This review aims to provide up-to-date information on crosstalk between these reactive species families and peroxisomes, as well as on their cellular environment in light of the well-recognized signaling properties of H2O2, NO and H2S.
INTRODUCTION
For many years, peroxisomes in higher plants have been given different names, such as glyoxysomes during seed germination and leaf senescence, as well as leaf, root and fruit peroxisomes according to their presence in different organs and at different physiological stages (Tolbert and Essner, 1981;Palma et al., 2018). This is explained by the presence of metabolic pathways which appear to be specific to each type of peroxisome. However, peroxisomes, which share a number of metabolites and enzymes common to all types of peroxisome, is now the preferred term regardless of their specific metabolic characteristics (Pracharoenwattana and Smith, 2008). The most noteworthy metabolites and enzymes include H 2 O 2 and catalase, which are directly involved in the metabolism of reactive oxygen species (ROS) (Su et al., 2018;Sousa et al., 2019).
Peroxisomes have a simple morphological constitution composed of a single membrane surrounding an amorphous matrix. Over the last 30 years, an increasing number of new and often unexpected components and processes in these organelles have been identified del Río et al., 1992;Corpas et al., 1994, Barroso et al., 1999Reumann et al., 2009;Clastre et al., 2011;Simkin et al., 2011;Chowdhary et al., 2012;Guirimand et al., 2012;Oikawa et al., 2015;Reumann and Bartel, 2016;Kao et al., 2018;Pan et al., 2018Pan et al., , 2020Borek et al., 2019), indicating that the plant peroxisomal metabolism and consequently peroxisomal enzymatic and non-enzymatic components are more diverse than previously predicted. The diverse complementary range of experimental approaches used to identify these new peroxisomal constituents includes: (i) the biochemical, proteomic and molecular analysis of purified peroxisomes combined with bioinformatics methodologies and (ii) cell biology studies of features such as immune localization with the aid of electron microscopy and specific fluorescent probes with appropriated controls. Although the model plant Arabidopsis thaliana has increased our knowledge of plant peroxisomes, it should be pointed out that studies of peroxisomes from other plant species have been essential, as the peroxisomal metabolism can be modulated depending on the plant organ, development time and plant species involved. Therefore, this review principally aims to provide an update of research on the metabolism of reactive species associated with oxygen, nitrogen and, more recently, sulfur, as well as to outline new challenges and possible future research perspectives regarding crosstalk between peroxisomes and other subcellular compartments such as oil bodies, mitochondria and plastids which are closely related both biochemically and structurally (Palma et al., 2006;Oikawa et al., 2019). Information on plant peroxisomes could also be useful in relation to peroxisome research into other organisms and vice versa.
PEROXISOMAL ROS METABOLISM
Reactive oxygen species (ROS) are produced by a series of singleelectron reductions in molecular oxygen which sequentially form superoxide (O 2 •− ), hydrogen peroxide (H 2 O 2 ) and hydroxyl (HO • ) radicals and ultimately ending in water (Figure 1). It is worth noting that the term peroxisomes, formerly known as microbodies, originates from their high H 2 O 2 content (De Duve and Baudhuin, 1966;Corpas, 2015). Plant peroxisomes contain a significant number of enzymatic systems capable of generating H 2 O 2 such as glycolate oxidase (GOX), acyl-CoA oxidase (AOX), urate oxidase (UO), polyamine oxidase, copper amine oxidase (CuAO), sulfite oxidase (SO), sarcosine oxidase (SOX), or superoxide dismutase (SOD) (Hauck et al., 2014;Corpas et al., 2017a and references therein). These H 2 O 2 -generating enzymes are involved in multiple biochemical pathways which are essential not only for the endogenous metabolism of plant peroxisomes but also for their interactions with other subcellular compartments such as plastids, mitochondria, cytosols, oil bodies and nuclei. In these subcellular interconnections, H 2 O 2 itself plays a highly important role as a signal molecule in crosstalk between organelles in order to coordinate cell function.
Photorespiration has been estimated to be responsible for 70% of total H 2 O 2 generated mainly from peroxisomal GOX in photosynthetic tissues (Noctor et al., 2002). Zhang et al. (2016) have described an elegant dynamic physical GOX-catalase association-dissociation mechanism that fine-tunes peroxisomal H 2 O 2 in rice plants. Although peroxisomal H 2 O 2 is kept under control when GOX and catalase are associated, under stress conditions and when mediated by salicylic acid (SA), this complex GOX-catalase dissociation mechanism inhibits catalase FIGURE 1 | Reactive oxygen species (ROS) produced from a sequential one-electron reduction from oxygen.
activity, leading to an increase in cellular H 2 O 2 which acts as a signaling molecule (Zhang et al., 2016;Kohli et al., 2019). Another sophisticated mechanism, involving the interaction of the γb protein from the barley stripe mosaic virus with GOX, has been reported to inhibit GOX and to facilitate infection with the virus (Yang et al., 2018). More recently, Yamauchi et al. (2019) observed a connection between the H 2 O 2 -generating GOX and catalase, which is required in the stomatal movement. Thus, when there is an increase of oxidized peroxisomes they were removed by pexophagy allowing an increase in H 2 O 2 in guard cells which mediated the stomatal closure. This mechanism of ROS homeostasis in guard cells seems to be relevant in response to environmental changes. On the other hand, the new peroxisomal small heat shock protein Hsp17.6CII, capable of increasing catalase activity especially under stress conditions, has been reported to be present in Arabidopsis plants (Li et al., 2017).
Acyl-CoA oxidase is another key peroxisomal H 2 O 2generating enzyme involved in fatty acid β-oxidation which, in collaboration with lipid bodies, enables triacylglyceride mobilization especially during seed germination and is also involved in the synthesis of signal molecules such as jasmonic acid (Baker et al., 2006;Chen et al., 2019b;Wang X. et al., 2019;Xin et al., 2019). However, under stress conditions such as salinity, ROS generated by peroxisomal fatty acid β-oxidation have a negative impact and contribute to oxidative damage .
Polyamines such as putrescine, spermidine and spermine are well known to be involved in multiple physiological processes, as well as mechanisms of response to various stress conditions (Wuddineh et al., 2018;Chen et al., 2019a;Wang W. et al., 2019). Several enzymes involved in the catabolism of polyamine, including H 2 O 2 -producing polyamine oxidase (PAO) and copper amino oxidase (CuAO), have been reported to be present in plant peroxisomes (Moschou et al., 2008;Kusano et al., 2015). These enzymes are also involved in the γ-aminobutyric acid (GABA) biosynthesis signaling pathway (Zarei et al., 2015;Corpas et al., 2009b).
Although catalase is the principal antioxidant enzyme in the matrix of all types of peroxisome (Mhamdi et al., 2010(Mhamdi et al., , 2012Palma et al., 2020 and references therein), other enzymatic antioxidants are present in both the matrix and the membrane. It is also important to highlight the role of SOD isozymes, which differ according to peroxisomal origin . Thus, peroxisomes of watermelon cotyledons have two SOD isoenzymes, a CuZn-SOD located in the matrix and a Mn-SOD that is bound to the membrane Rodríguez-Serrano et al., 2007); pea leaf peroxisomes have a Mn-SOD present in the matrix; sunflower cotyledon peroxisomes have only a CuZn-SOD which is also located in the matrix ; carnation petal and pepper fruit peroxisomes have a Mn-and an Fe-SOD (Droillard and Paulin, 1990;Palma et al., 2018); and olive fruits peroxisomes contain four SOD isozymes, an Fe-SOD, two CuZn-SOD and a Mn-SOD (López-Huertas and del Río, 2014). Therefore, it could be hypothesized that the presence of two or more types of SOD in peroxisomes must have some physiological advantages. Thus, one of the SOD isozymes could be constitutive while the other one could be inducible under environmental or physiological stimuli such as seedling development, leaf senescence or fruit ripening.
In addition, it is worth noting the role of ascorbate-glutathione cycle components, including ascorbate peroxidase (APX), monodehydroascorbate reductase (MDAR), dehydroasrcorbate reductase (DAR) and glutathione reductase (GR) (Jiménez et al., 1998;Romero-Puertas et al., 2006;López-Huertas and del Río, 2014;. While MDAR is present in both matrix and membrane (Leterrier et al., 2005;Lisenbee et al., 2005;Eastmond, 2007), APX is exclusively located in the membrane (Corpas et al., 1994;Yamaguchi et al., 1995;Bunkelmann and Trelease, 1996). With its high affinity for H 2 O 2 (low Km value around 74 µM), membrane-bound APX appears to have fine-tuned control of H 2 O 2 (Ishikawa et al., 1998) as compared to catalase, which, with a Km value in the mM range, is less efficient at low concentrations of H 2 O 2 (Huang et al., 1983;Mhamdi et al., 2010). The Km values for plant catalase are reported to vary quite considerably, with, for example, a Km of 50 mM in Beta vulgaris (Dinçer and Aydemir, 2001), 100 mM in rice (Ray et al., 2012) and 190 mM in pea (del Río et al., 1977). Peroxisomal APX appears to be critical in a diverse range of processes such as seedling development and leaf senescence (Ribeiro et al., 2017). To maintain the ascorbate-glutathione cycle at the GR level, NADPH needs to be supplied by NADP-dependent endogenous dehydrogenases including glucose-6-phosphate dehydrogenase (G6PDH), 6-phosphogluconate dehydrogenase (6PGDH) and isocitrate dehydrogenase (NADP-ICDH) (Leterrier et al., 2016;Corpas and Barroso, 2018b and references therein). In addition, have reported the presence of a protein immunologically related to plant peroxiredoxins, whose expression is differentially modulated under oxidative stresses such as those induced by CdCl 2 and the herbicide 2,4-dichlorophenoxyacetic acid (2,4-D); however, further research is necessary to clarify this phenomenon. Figure 2 shows a working model of the ROS metabolism and its interaction with other reactive species, including NO and H 2 S, which modulate the activity of peroxisomal enzymes through posttranslational modifications (PTMs), events which will be further discussed below.
Given the capacity of ROS to mediate several PTMs, particularly carbonylation and S-sulfenylation, certain amino acid residues, especially arginine, lysine, threonine and proline, are carbonylated, which affects target protein function in many cases (Debska et al., 2012;Lounifi et al., 2013). Several studies have identified peroxisomal proteins, such as catalase, malate synthase and the fatty acid β-oxidation multifunctional protein AIM1, which undergo carbonylation (Nguyen and Donaldson, 2005;Anand et al., 2009;Mano et al., 2014;Rodríguez-Ruiz et al., 2019). On the other hand, H 2 O 2 can oxidize specific protein cysteine thiols to sulfenic acid (SOH), a process known as S-sulfenylation, which usually results in enzymatic inactivation. Using proteomic techniques, approximately 2% of peroxisomal proteins have been reported to be susceptible to S-sulfenylation (Akter et al., 2017;Huang et al., 2019). This PTM has been observed to occur with respect to fatty acid β-oxidation acyl-coenzyme A oxidase 1, the multifunctional proteins MFP2, and AIM1, as well as amine oxidase, phosphomevalonate kinase, MDAR and NADP-ICDH. Table 1 shows a summary of peroxisomal enzymes targeted by carbonylation and S-sulfenylation, as well as other PTMs mediated by RNS and RSS, a subject which will be discussed below.
Given growing awareness of the important role of ROS peroxisomal metabolism in combating biotic stress, the expression of genes encoding for peroxisomal proteins involved in their biogenesis, fatty acid catabolism and the H 2 O 2 -generating glyoxylate cycle have been reported to increase during interactions between the pathogen Sclerotinia sclerotiorum and rapeseed (Brassica napus), thus facilitating pathogen cell wall degradation and metabolism detoxification (Chittem et al., 2020). On the other hand, using the Arabidopsis nca1 mutant with no catalase activity 1, containing residual activity of the three catalase isozymes, Hackenberg et al. (2013) identified a link between catalase and ROS production as autophagy-dependent cell death progresses. Table 2 shows some functional implications of peroxisomal H 2 O 2 and other signal molecules generated in this organelle. •− ) which is dismutated to H 2 O 2 by superoxide dismutase (SOD). All three SOD types have been described in plant peroxisomes from different origin, CuZ-SOD, Mn-SOD, and Fe-SOD. The H 2 O 2 pool is mainly decomposed by catalase (CAT) but also by the membrane-bound ascorbate peroxidase (APX). An L-arginine (L-Arg) and Ca 2+ dependent NOS-like activity generates NO which can react chemically with O 2 to produce peroxynitrite (ONOO -), a nitrating molecule that facilitates PTMs such as tyrosine nitration. NO can also interact with reduced glutathione (GSH) to form S-nitrosoglutathione (GSNO), a NO donor which mediates S-nitrosation. GSH is regenerated by glutathione reductase (GR) which requires NADPH supplied by several NADPH-generating enzymes (NADPH-ICDH, G6PDH, and 6PGDH). Uric acid is a ONOOscavenger, this being a mechanism of peroxisomal auto-regulation. With all these components, and according to reported data, the peroxisomal targets of NO-derived PTMs identified so far are CAT, CuZn-SOD, and monodehydroascorbate reductase (MDAR) which can undergo an inhibitory effect either by nitration or S-nitrosation. Additionally, CAT and GOX can be inhibited by hydrogen sulfide (H 2 S), and CAT is also inhibited by carbonylation. The H 2 O 2 -generating sulfite oxidase (SO) converts sulfite (SO 3 2-) to sulfate (SO 4 2-), which is a mechanism of protection because sulfite inhibits catalase activity. Red line denotes inhibition effect.
The generation of singlet oxygen ( 1 O 2 ) has always been associated with chloroplasts, particularly in photosystem II, responsible for various types of photo-damage which triggers distinct cellular responses (Wagner et al., 2004;Rosenwasser et al., 2011;Chen and Fluhr, 2018;Dogra et al., 2018). Using the green fluorescence probe to detect 1 O 2 , peroxisomes, mitochondria and nuclei have been shown to be either the origin or target of 1 O 2 , suggesting that this ROS is generated in a light-independent manner (Mor et al., 2014). These findings open up new questions about the importance of 1 O 2 in the mechanism of response to plant stress in which several subcellular compartments including peroxisomes are involved.
PEROXISOMAL REACTIVE NITROGEN SPECIES (RNS)
Nitric oxide (NO) metabolism has a significant impact on cellular metabolisms due to its involvement in the important plant physiological processes of seed and pollen germination, root development, stomatal closure, senescence and fruit ripening, as well as in the mechanism of response to many environmental stresses including salinity, drought, heavy metals and extreme temperature (Neill et al., 2008;León et al., 2014;Begara-Morales et al., 2018;Kolbert et al., 2019;Wei et al., 2020). NO belongs to a family of related molecules called reactive nitrogen species (RNS), with peroxynitrite (ONOO − ) and S-nitrosogluthione (GSNO) being the most studied. Using various experimental approaches including electron paramagnetic resonance (EPR) spectroscopy, as well as biochemical and cellular biology, some RNS including NO, ONOO − and GSNO have been detected in plant peroxisomes (Barroso et al., 2013;Corpas and Barroso, 2014b;Corpas et al., 2019). Identification of peroxisomal proteins undergoing PTMs mediated by these NO-derived species is strong evidence of an active RNS metabolism in peroxisomes. Figures 3A-H shows in vivo images of NO and ONOO − in Arabidopsis guard cell peroxisomes detected by confocal laser scanning microscopy (CLSM) and specific fluorescent probes. ONOO − results from a reaction between NO with O 2 •− , considered one of the fastest chemical reactions with a rate constant (k) of 1.9 × 10 10 M −1 s −1 (Kissner et al., 1997). ONOO − , a strong oxidant and nitrating molecule involved in protein tyrosine nitration (NO 2 -Tyr), modifies protein function, mostly through inhibition (Corpas et al., 2009a;Mata-Pérez et al., 2016). This NO-derived PTM involves the covalent oxidative addition of a nitro group (-NO 2 ) to tyrosine residues, a highly selective process which depends on factors such as the protein environment of the Tyr and the nitration mechanism (Bartesaghi and Radi, 2018). Table 1 shows some nitrated proteins identified in plant peroxisomes and how their function is affected. Interestingly, some of the proteins affected are directly involved in the ROS metabolism, indicating a close metabolic interconnection between both families of reactive species.
The antioxidant glutathione (GSH), a tripeptide (γ-Glu-Cys-Gly), undergoes S-nitrosation in order to generate GSNO, a low-molecular-weight NO reservoir, through a covalent addition of NO to the thiol group of Cys residues in order to form S-nitrosothiol (SNO) . GSNO is a key molecule given its dynamic interaction with free cysteines, GSH and proteins through processes such as S-nitrosation, S-transnitrosation and S-glutathionylation (Broniowska et al., 2013;Corpas et al., 2013a,b). GSNO is enzymatically decomposed by GSNO reductase (GSNOR; Leterrier et al., 2011), an enzyme susceptible to S-nitrosation and consequently inhibition (Guerra et al., 2016). An increase in Tyr nitration, an irreversible process, is usually associated with nitro-oxidative stress; however, protein S-nitrosation, a reversible process, is a regulatory protein mechanism that occurs under physiological and stress conditions. Table 1 shows some peroxisomal proteins targeted Abiotic stress tolerance Shelp and Zarei, 2017 by S-nitrosation, as well as proteins involved in ROS metabolism which are targeted by these NO-mediated PTMs.
The number of peroxisomal proteins targeted by NOmediated PTMs is growing continuously. Using the biotin-switch technique and liquid chromatography/mass spectrometry/mass spectrometry (LC-MS/MS), several more S-nitrosated peroxisomal proteins have been identified during adventitious root growth induced by treatment with NO (Niu et al., 2019). These proteins include the peroxisomal LON2 protease, which is necessary for matrix protein import into peroxisomes (Lingard and Bartel, 2009); isocitrate lyase (ICL), involved in the glyoxylate cycle; and the multifunctional AIM1-like isoform, involved in fatty acid β-oxidation.
However, the source of enzymatic NO, as yet unelucidated, is currently the most controversial aspect of NO metabolism in higher plants (Kolbert et al., 2019). Two main candidates have been proposed: nitrate reductase (NR) (Mohn et al., 2019) and L-arginine-dependent NO synthase-like activity . Although no evidence of NR has been found in plant peroxisomes, NO synthase-like activity has been found and characterized in peroxisomes purified from pea leaves (Barroso et al., 1999). Though as yet unidentified, this protein is called NOS-like activity, as peroxisomal NO generation requires NOS proteins similar to those found in animals, including L-arginine, NADPH, FMN, FAD, tetrahydrobiopterin, calcium, and calmodulin (Corpas and Barroso, 2017b;Corpas et al., 2019). The protein responsible for NO generation is imported by a type 2 peroxisomal targeting signal involving a process dependent on calmodulin and calcium (Corpas andBarroso, 2014a, 2018a).
Peroxisomal NO metabolism is involved in processes such as pollen tube germination (Prado et al., 2004), lateral root formation (Schlicht et al., 2013), and leaf senescence , as well as in responses to environmental and heavy metal stresses such as salinity (Corpas et al., 2009b), lead , and cadmium (Corpas and Barroso, 2014b;Piacentini et al., 2020).
Different molecules and enzymes, such as GSH (Müller et al., 2004), glutathione reductase (Romero-Puertas et al., 2006), and sulfite oxidase (Nowak et al., 2004;Hänsch and Mendel, 2005), involved in sulfur metabolism, are present in plant peroxisomes. Sulfite oxidase (SO) catalyzes the conversion of sulfite to sulfate by producing H 2 O 2 . The functional relevance of this enzyme is that it can protect catalase activity since sulfite, at low concentration, has the capacity to inhibit catalase activity (Veljovic-Jovanovic et al., 1998). Nevertheless, despite the greater importance attributed to peroxisomal SO in a recent study, mitochondrial SO in animal cells has the capacity to generate NO from nitrite (Bender et al., 2019), while NO enzymatic generation from SO in plant peroxisomes remains to be proven. An earlier study confirmed the important role played by the peroxisomal RSS metabolism (Corpas and Barroso, 2015).
H 2 S has recently been proven to be present in plant peroxisomes (Corpas et al., 2019a). Figures 4A-G shows representative images of H 2 S in peroxisomes from the root tips and guard cells of Arabidopsis seedlings detected by in vivo CLSM and a specific fluorescent probe. Using proteomic techniques, some peroxisomal enzymes have been identified as targets of persulfidation (Aroca et al., 2015(Aroca et al., , 2017. On the other hand, in vitro analysis shows that catalase activity from Arabidopsis and sweet pepper fruits is inhibited in the presence of H 2 S (Corpas et al., 2019a). Although, to our knowledge, the enzymatic source of peroxisomal H 2 S remains unknown, previous studies have proposed some potential candidates. For example, catalase, which functions as a sulfide oxidase or sulfur reductase, is capable of oxidizing or generating H 2 S (Olson et al., 2017). SOD has also been reported to have the capacity to catalyze the reaction between O 2 and H 2 S to generate persulfide (Olson et al., 2018). In a previous study by Corpas and Barroso (2015), the presence of these enzymatic and non-enzymatic components in plant peroxisomes indicates that, in addition to ROS and RNS, these organelles also have an active RSS metabolism.
CROSSTALK BETWEEN PEROXISOMAL REACTIVE SPECIES
Functional interactions and inter-regulation through PTMs in these families of reactive species are shown in Figure 2. In this working model, under physiological conditions, catalase, the main antioxidant enzyme, regulates levels of H 2 O 2 generated by different pathways, principally photorespiratory glycolate oxidase (GOX) (Noctor et al., 2002). On the other hand, peroxisomal xanthine oxidoreductase (XOR) activity involved in purine catabolism generates uric acid, with the concomitant formation of the O 2 •− (Corpas et al., 1997(Corpas et al., , 2008Zarepour et al., 2010), which, in turn, is dismutated to H 2 O 2 by SOD. The pool of H 2 O 2 is mainly decomposed by catalase (CAT) and also by membranebound ascorbate peroxidase (APX). L-Arg-dependent NOS-like activity generates NO which chemically reacts with O 2 •− to produce peroxynitrite (ONOO − ), a nitrating molecule that facilitates PTMs such as tyrosine nitration. NO also interacts with reduced glutathione (GSH) to form S-nitrosoglutathione (GSNO), a NO donor that mediates S-nitrosation. Uric acid is a physiological ONOO − scavenger (Alamillo and García-Olmedo, 2001) involved in endogenous peroxisomal auto-regulation. Thus, the peroxisomal enzymes targeted by NO-derived PTMs, catalase (CAT), CuZn-SOD, and monodehydroascorbate reductase (MDAR), are inhibited by nitration and S-nitrosation. Both CAT, and GOX are inhibited by H 2 S; the former is also inhibited by carbonylation when H 2 O 2 is overproduced. In addition, H 2 O 2 -generating sulfite oxidase (SO) is involved in the conversion of sulfite to sulfate which, given sulfite's ability to inhibit SO, has been reported to be a catalase protection mechanism. These interconnections highlight the biochemical complexity of this self-regulated plant peroxisome network, in which the antioxidant catalase is one of the most regulated peroxisomal enzymes (Palma et al., 2020).
CONCLUSION
Much of our knowledge of reactive species metabolism in plant peroxisomes is now well established. The three molecular families ROS, RNS, and RSS are present in plant peroxisomes, which are considered to be potential producers of reactive species and to play an important role in the cell signaling network. However, our limited knowledge of reactive species families needs to be expanded by identifying new peroxisomal protein targets. We also need to determine the effect of the different PTMs, carbonylation, S-sulfenylation, S-nitrosation, tyrosine nitration, and persulfidation, on target protein function and peroxisomal metabolism. In addition, interactions with other subcellular compartments which share biochemical pathways such as photorespiration, fatty acid β-oxidation, isoprenoid biosynthesis and purine and polyamine metabolism (Clastre et al., 2011;Guirimand et al., 2012;Corpas et al., 2019b) should be investigated. Similarly, the relationship between reactive species and complex peroxisomal biogenesis, division and matrix/membrane protein import mechanisms (Reumann and Bartel, 2016;Kao et al., 2018) has been underexplored (López-Huertas et al., 2000). Further research should also be carried out to identify the proteins responsible for endogenous peroxisomal generation of NO and H 2 S. This would increase our knowledge of how organelle biochemistry is modulated within the framework of the whole cell metabolism. This research could lead to biotechnological applications given the important role of peroxisomes in many physiological processes and in responses to biotic and abiotic stresses. Furthermore, in addition to harboring reactive species, with their known signaling properties, peroxisomes are a source of other signaling molecules such as jasmonic acid and γ-aminibutic acid (GABA), which extends the functional role of plant peroxisomes. Table 2 shows signaling molecules generated in the plant peroxisomal metabolism and some examples of their role in various plant processes.
AUTHOR CONTRIBUTIONS
Authors have made a collaborative, direct and intellectual contribution to the work, and have approved it for publication.
FUNDING
This work was supported by the European Regional Development Fund co-financed grant from the Ministry of Economy and Competitiveness (AGL2015-65104-P and PID2019-103924GB-I00), the Plan Andaluz de Investigación, Desarrollo e Innovación (P18-FR-1359), and the Junta de Andalucía (group BIO192), Spain. | 5,829.6 | 2020-07-03T00:00:00.000 | [
"Biology"
] |
Knowing Customers Better: An Experimentation of Twit Marketing in the e-Commerce Industry
—Internet gives incredible opportunities for companies to learn about their consumers and provides different channels for marketing. Otlobli.com website (an e-commerce company established in Saudi Arabia), uses different channels, such as: Google AdWords, newspaper web-site banners, official thread on a famous online forum and, Twitter. A marketing tool has been developed to enhance the effectiveness of marketing on Twitter. This research investigates the impact of the developed system. Website access from these various channels was examined, in addition to the number of visitors who have completed an order. Results showed that the tool caused around 49% increase in the number of visits from Twitter, and comparatively, a good percentage of them completed an order.
INTRODUCTION
The fundamental thing marketing professionals want is to learn about customers.Internet gives incredible opportunities for companies to learn about their consumers.In the knowledge economy, knowledge is increasingly considered a critical factor for organization's success and a source of its competitive advantage and value creation [1], [2].The development and widespread use of information technologies are fostering the revolution of means used to access, and share information.In particular, internet technologies are literally facilitating individual and organizational access to information they need and utilize these information as needed.This new economy has introduced a new lexicon in which knowledge capital, intellectual capital, learning organizations, intangible assets, and human capital describe new forms of economic value [3].The potential usefulness of information technologies is well recognized.In fact, these technologies are considered critical part of modern life.Their importance has rapidly increased with time on a personal and organizational levels, especially with the development of the Internet and electronic commerce.Now a days, the role of IT is more essential rather that supportive.It helps in building organizational strategies by providing new market opportunities, products, and customers [4].
Social media have revolutionized the 21st century.The proliferation of social media has exceeded expectations.It is globally available on numerous devices, including mobile devices, which assist users to access these services any time.With social networking sites, such as Twitter and Facebook, people are connected with each other and can share and exchange messages, media, and news easily.Social media have become essential tools for individuals, companies, and governments.The influence of social network sites is critical and far-reaching, affecting personal lives, governments, and business, particularly in terms of marketing strategies, sales, and new opportunities for existing and start-ups.Lately, marketing is more significant than ever, with many channels available that can be used to reach customers and optimally market goods and services.Marketing has been performed directly or indirectly via multiple traditional channels including TV, newspapers, banners, or other print media.However, the introduction of the World Wide Web has changed the concept of marketing, with new channels that supersede traditional ones.
This study aims to develop and experiment a tool for marketing on Twitter for one of the e-commerce sites, Otlobli.com, a company in Saudi Arabia that handles the purchasing and payment processes of online products and ensures the delivery of these items to local addresses.
II. TWITTER FOR MARKETING
Social media has influenced advertisements and marketing strategies.Recently, there has been an interest in the used of social media in marketing.Several companies are already using social networking sites to support the creation of brand communities [5] or for marketing research in the context of netnography [6].
Twitter, which was launched in 2006, enables users to send and read short messages limited to 140 characters, known as "tweets".For some authors referred to Twitter as an information sharing tool rather than a social network platform [7], [8], [9].Regardless of its classification, Twitter has grown exponentially, reaching 200 million users and recording over 30 billion messages in only five years [10].
Twitter is one of the most known and accessed social networking site in Saudi Arabia.Twitter penetration keeps breaking records, placing Saudi Arabia on top Twitter user penetration per internet capita, increasing its user base by 45% in 2013 with more than 5 million active users, 73% of them access Twitter via mobile (http://www.thesocialclinic.com).Most users of Twitter share or report news, exchange information, discuss different opinions and political and social interests, and campaigning for or against certain people, companies, or governments [11].
In addition of the role Twitter plays at different levels; for instance, at the political level as in the Egyptian revolution in 2011 [12] on one hand, and on the business level as in word of mouth marketing on the other hand [13], Twitter can be considered an effective platform for mar-
SHORT PAPER KNOWING CUSTOMERS BETTER: AN EXPERIMENTATION OF TWIT MARKETING IN THE E-COMMERCE INDUSTRY
keting and public relations.Marketing strategies on Twitter can be, interactive or reactive.The former is using Twitter "to provide a highly interactive one-to-many information channel, using, and a combination of retweets, hyperlinks, and hashtags to promote positive messages, especially by independent influential individuals" [14].Where the latter is "using Twitter as a service recovery channel to respond to customer complaints -both those made directly to the organization, or those discovered by monitoring the Twitter feed" [13].Twitter can also provide easy access to information for those to who it is most relevant or interesting, by pushing users to an internal web site on one hand, and companies can listen to/influence consumers' opinions" [15].
Otlobli (otlobli.com) is a small Saudi online company that handles the purchasing, payment and delivery processes of online products from other e-commerce sites and ensures the delivery of purchased items to local addresses.They use different channels to market their services many of which are performed manually.Some examples of these channels are Twitter, Google AdWords, newspaper website banners, and advertisements in online forums.
The company never tried a semi-automated mechanism with Twitter to market its services.Therefore, we developed and experimented a Twit Marketing tool, which is discussed in section III, for advertising Otlobli and we compared it to the other techniques.Results are presented in section V.
III. TOOL DEVELOPMENT FOR TWIT MARKETING
Lots of questions being asked by people on Twitter either were not receiving answers, or were receiving useless answers.Businesses are having a hard time finding ways to interact with Twitter users.There are various tools available in the market, such as: InboxQ (inboxq.com),Twitter Keyword Tracker (twtrland.com),and Tweetbeep (tweetbeep.com).However, most of the available solutions are limited to basic keyword searches; and none of these tools provide an automatic replies to the tweets that found to be related to a specific topic.A tool that helps increase the likelihood that information seekers will get good answers to their questions has been developed.
The web-based tool can be used to collect everything about a specific topic based on the keywords related to the topic (e.g.product and organization).We can then reply automatically to all tweets collected with a pre-defined reply.This tweet can also be sent to a group of people who are potentially interested in buying online or looking for some help in this matter, system architecture is shown in Figure I.In particular, the web-based tool is used in two steps.First, creating categories and defining keywords that can be used to collect business related tweets.Second, replying to the collected tweets in which multiple tweets can be selected from the list and a single reply can be sent from company's Twitter accounts.Automatic replies has been defined to be sent immediately to tweets collected under a specific category.The tool also provides a statistical report on all tweets, including the retweets, accounts been tweeted …etc.
One important consideration in the development of the tool is the ethical issue.The tool can be used to help information seekers.But also can be misused to harm people or companies by sending messages to people who do not wish to receive them.To avoid any ethical violation, we manually made sure that tweets sent by the system reaches only people who are interested in such information and avoid any random push of messages.
IV. METHODOLOGY
As mentioned previously, Otlobli.commarketing partner used to send tweets manually to potential customers who ask about buying online or related information.We conducted an experimentation at the company to study the impact of Twit Marketing on website visitors and number of orders made.
The research have gone through two main stages.First, in collaboration with Otlobli management, we manually listed 35 keywords or hashtags that indicates potential customers.For instance, one of the detected hashtags was related to a negative feedback regarding another online company that provides similar service.Second, the monitoring period of three weeks for collecting interesting tweets and subsequently sending a 113-character tweet that summarizes Otlobli's main service and the link to the web site." !"#$%&' ()*+, -./ (01)23)4 526721)4' 894*1)4 :; <%=>2; ?) @=$AB !"# 4 !"#"$ .%&'()*+,#-" www.otlobli.com The accounts are configured to send one tweet every five minutes per account, enabling us to send 100 tweets Later on, Otlobli management provided detailed statistics about the number of visits, including statistics on who completed an order, allowing us to present accurate results.In addition, we got analytics distributed on the channels used for marketing, for example, Google Adwords, normal banners on certain websites, and publishing official advertisement threads on an online forum.These statistics can be found via Google Analytics, a comprehensive, accurate, and trusted third-party tool.
V. FINDINGS
An experiment has been conducted to measure the effectiveness of marketing via Twitter and it has gone through two stages, hunting potential customers' tweets, and then replying to those tweets and measure the impact on the number of visitors and completed orders.During the three-week stage, many Arabic tweets (n = 567) were captured, examples are shown in Table I.The majority of these tweets (n = 432) were related to one hashtag in which users expressed their dissatisfaction of dealing with a competitor company providing similar services.Obviously, this number of related tweets is difficult to capture manually and will require a lot of time and effort.
As stated previously, different channels are used to market the Otlobli.comwebsite, these are: Google Ad-Words, newspaper website banners, official thread on a famous online forum and Twitter.However, to investigate the accurate value of the developed system, we differentiate between the manual/traditional marketing, and the semi-automated marketing using the developed tool.Website access from these various channels was examined, in addition to the number of visitors who have completed an order.Data has been gathered captured over a four-day period from Google Analytics are presented in Table III.An explanation of the metrics is provided in Table II.
Manual/traditional marketing on Twitter is conducted through a company that specializes in marketing on Twitter.The company undertakes several activities, such as replying to all questions forwarded to the Otlobli account on Twitter, following new accounts to indirectly inform them about the Otlobli account, promoting tweets in active hashtags, tweeting daily using the official Otlobli accounts on Twitter, and re-tweeting some Otlobli account tweets.The average of visits that used to come from Twitter, using the traditional marketing method, were around the 550 visits for a similar four-day period.
After the start of the experimentation (replying to the collected tweets), the number of visits to the website raised to 822 visits in the four-day period.We can therefore conclude that the additional visits were because of the replies sent automatically through the developed system.
VI. CONCLUSION
Google AdWords is clearly the best marketing channel based on all the metrics (visits and orders completed).It was the cause of most of the visits (79%) to Otlobli.com website during the four-day period.However, only 2.91% of those visitors made completed order, see Table IV.On the other hand, Twitter followed Google AdWords in rank.Twitter caused 822 visits, (17%) of total visits.More importantly, the percentage of visitors who actually placed orders is not substantial between Google Ad-Words and Twitter marketing channels.Since 2.91% of visitors through Google AdWords placed actual orders, 1.88% of visitors through Twitter submitted complete orders.The number of visits from Twitter before using the marketing tool was about 550 visits and the tool caused the increase of about 49% getting 822 visits in total.
TABLE I .
SAMPLE OF CAPTURED TWEETSPercentage of visitors who submit and pay for a specific order on Otlobli.com from the total number of visits. | 2,807 | 2014-06-08T00:00:00.000 | [
"Business",
"Computer Science"
] |
Surrogate Modelling for Oxygen Uptake Prediction Using LSTM Neural Network
Oxygen uptake (V˙O2) is an important metric in any exercise test including walking and running. It can be measured using portable spirometers or metabolic analyzers. Those devices are, however, not suitable for constant use by consumers due to their costs, difficulty of operation and their intervening in the physical integrity of their users. Therefore, it is important to develop approaches for the indirect estimation of V˙O2-based measurements of motion parameters, heart rate data and application-specific measurements from consumer-grade sensors. Typically, these approaches are based on linear regression models or neural networks. This study investigates how motion data contribute to V˙O2 estimation accuracy during unconstrained running and walking. The results suggest that a long short term memory (LSTM) neural network can predict oxygen consumption with an accuracy of 2.49 mL/min/kg (95% limits of agreement) based only on speed, speed change, cadence and vertical oscillation measurements from an inertial navigation system combined with a Global Positioning System (INS/GPS) device developed by our group, worn on the torso. Combining motion data and heart rate data can significantly improve the V˙O2 estimation resulting in approximately 1.7–1.9 times smaller prediction errors than using only motion or heart rate data.
Introduction
Oxygen consumption (VO 2 ) during unconstrained running can be measured directly by portable spirometers or metabolic analyzers. However, using these devices on a constant basis is not convenient and often requires trained personnel. Furthermore, the price range of portable spirometers can be in tens of thousands of euros, causing additional operational costs. Therefore, estimatingVO 2 based on measurements from consumer-grade sensors and without reliance on direct measurements would be highly useful for different performance assessment applications and enable their wide-range use.
Indirect estimation ofVO 2 from the observation of other variables (using surrogate variables) is a growing field of research thanks to the development of small wearable sensors and machine learning algorithms. Typically, surrogate modeling of instantaneous oxygen uptake or its steady value is based on heart rate (HR) measurements [1]. However, there are factors that can affect the relationship between heart rate and oxygen uptake. Some of these include hydration status, exercise duration, medication, altitude, state of training, environmental conditions, and time of day [2]. Therefore, some additional input features can be considered, and their choice depends on specific application. For example, in running it could be breathing frequency, speed, speed variation and cadence that can be computed by wearable devices during outdoor exercise [3]. For cycling, in addition to HR and breathing frequency, the input features can include mechanical power and pedaling cadence that can be accurately measured directly on cycling ergometers [4][5][6].
Many previous studies used linear regressions forVO 2 prediction [1,3,7,8]. However, this approach has limitations for very low and very high intensity exercises when the HR vs.VO 2 relationship is significantly non-linear. Furthermore, heart rate is affected by factors such as day-to-day variability, age, sex, fitness, exercise modality, and environmental conditions [3]. Due to a reduced number of parameters, analytical models cannot adapt to changes in exercise conditions. To mitigate the effect of these variables, HR index (HRI) can be used as a surrogate forVO 2 prediction instead of HR [1,8]. HRI is equal to a given HR divided by resting HR and it is considered that it can potentially remove the need for individual calibration often required for tracking daily activity using HR [8].
Some commercial products, for example, the Suunto personal HR monitoring system, use HR measurements as a surrogate for the estimation ofVO 2 and energy expenditure [7]. In this approach the input features include R-wave-to-R-wave (R-R) heartbeat intervals, R-R-derived respiration rate, and the on-and-offVO 2 dynamics during various exercise conditions. Although the investigators acknowledge the limitations in the prediction accuracy when individual values for maximal HR andVO 2 are included, they give little information on the validity against pulmonary gas exchange values or correction factors to account for variation in these estimates.
In [9], machine learning was used to predictVO 2 during walking and with other daily activities using surrogate variables such as HR, breathing frequency, minute ventilation, hip acceleration and walking cadence. ECG and respiration band were integrated into a Hexoskin smart shirt. Walking cadence was computed based on hip acceleration using proprietary algorithm. From the HR data, a new variable was derived. The ∆HR was composed by the difference between the current HR value and the previous value by a 1 s lag operator, capturing dynamic changes in cardiac activity. TheVO 2 data were measured breath-by-breath by a portable metabolic system (K4b2, COSMED, Italy). The predicted oxygen uptakeVO 2 was obtained by a random forest regression based on these features streamed from wearable sensors throughout the day. The deployment of such nonintrusive technologies can help us to study relationships between patterns of daily physical activity and fitness markers. The prediction accuracy was 6.166 mL/min/kg (95% limits of agreement). ForVO 2 prediction, advanced machine learning techniques can be used, such as long-short term memory (LSTM) neural networks (e.g., in [4], a recurrent neural network with three LSTM layers is used). These networks can process entire sequences of data and consider the historical context of inputs. This property is important for applications such as oxygen uptake predictions. Therefore, LSTM networks are widely used in the field of exercise physiology to modelVO 2 dynamics across a variety of exercise conditions. They perform better than analytical models or linear regressions [4]. In [4], an LSTM neural network was trained from laboratory cycling data collected on electromagnetically-braked bicycle ergometer (Excalibur Sport, Lode) to predictVO 2 values from easy-to-obtain inputs, such as heart rate, mechanical power output, cadence, and respiratory frequency. Accuracy of workload was 2% for power between 100 and 1500 Watt. The authors compared the performance of their LSTM neural network to that of two analytical models: (a) a first order model assumingVO 2 dynamics to be linear and (b) the model from [10]. The root mean squared error (RMSE) of the LSTM was about 7.2% of theVO 2max . For the analytical models the respective RMSEs were between 10 and 15% of theVO 2max , depending on the test intensity. One drawback of neural networks is that model parameters often do not have direct physiological meaning and require big datasets for training.
Amelard et al. [5] investigated the temporal prediction ofVO 2 from wearable sensors during cycle ergometer exercise using a temporal convolutional neural network (TCN) or stacked LSTM. The four cardiorespiratory bio-signals (HR, HR reserve, breathing frequency, and minute ventilation) derived from a Hexoskin smart shirt (Hexoskin, Carre Technologies, Montreal, Canada). The shirt contained a textile electrocardiogram to measure HR and thoracic and abdominal respiration bands to obtain estimates of breathing frequency and minute ventilation (V E ) via respiratory inductance plethysmography. The V E was also calibrated by linear regression to the known V E measured throughout each protocol with the bi-directional turbine [5]. In addition to these four inputs, the work rate profiles were used as inputs into neural network. TheVO 2 data were measured breath-by-breath by a portable metabolic system (MetaMax 3B-R2, CORTEX Biophysik, Leipzig, Germany). TCN showed aVO 2 prediction accuracy of approximately 8% of theVO 2max (95% limits of agreement) when the input data history was 218 s.
In this study, we investigated how motion data can contribute toVO 2 prediction and if it can be sufficient for accurateVO 2 prediction. We also check whether the addition of heart rate as an input feature can improveVO 2 prediction accuracy. We have developed a data-driven surrogate modelling for accurate oxygen uptake prediction based on motion data measured by an inertial navigation system (INS) combined with a Global Positioning System (GPS) receiver [3] and/or heart rate using an LSTM network. TheVO 2 predictions are individualized, i.e., training and predictions are for the same person.
Our measurement setup was based on an INS/GPS device attached to the torso [3] and a heart rate monitor (HRM). We developed a system that can provide continuously different walking and running metrics. We have already used this system for indirect estimation of vertical ground reaction force (vGRF), ground contact time (GCT), and some other target features that can be directly measured only by instrumented insoles or by force plates [3,11]. The results showed that the LSTM network can accurately predict vGRF and GCT based on measurements of accelerations and angular velocities [11].
Considering the promises of LSTM networks, the purpose of this study was to predict the individual response ofVO 2 during unconstrained walking and running using measurement from the body-mounted motion sensor and/or HRM. One research goal was to study how motion data contribute toVO 2 prediction and if the motion data alone can achieve a prediction accuracy comparable to HRM based predictions. We also tried to select an optimal LSTM NN model: input features and length of input sequences, number of LSTM layers and hidden units. Agreement between the measured and predicted oxygen uptake is validated by Bland-Altman analysis by computing root mean square errors.
Experimental Data
Twelve recreational runners (age 30.9 ± 8.7 yrs, height 176.9 ± 8.9 cm, body mass 78.3 ± 13.0 kg, body mass index 24.9 ± 3.1 kg/m 2 , nine males) with varying fitness levels participated in the field tests on a level outdoor track. All participants were healthy and trained, according to their own statements, on average two times per week. The Ethics Committee of Tampere Region approved the study. The participants gave informed consent, and the research was conducted in accordance with the WMA Declaration of Helsinki [12]. After arriving at the track and providing consent, the participants were equipped with the chest mounted INS/GPS device, oxygen mask, gas analyzers and telemetry radios from Oxycon Mobile and, in some cases, a heart rate monitor's chest strap. The setup is shown in Figure 1. Participants were instructed to walk or run on the outdoor track at speeds ranging between 1.3 and 3.5 m/s. The speed was chosen subjectively based on the participant's own feeling of fast and slow pace.
Every test for each test person started from resting condition (sitting or standing still) for 3 min to measure oxygen consumption at rest. This allowed us to see the effect of exercise on oxygen consumption. The test continued with walking at two different speeds and running at two different speeds. Thus, data from four speeds and two different gaits were collected. The participants maintained each speed for 3 min so that the oxygen consumption had time to reach a steady state. In between each tested speed, the participants were standing still to allow oxygen uptake and heart rate to return to resting level. The total duration of tests for one participant was about 20 min. The data were collected in two measurement campaigns, on 24/25 May and 1/2 November 2022 in Jyväskylä, Finland. The field tests were carried out at different temperatures, ranging from +3 • C up to +22 • C, and different relative humidity levels, ranging from 21% to 100%.
INS/GPS datalogger
Oxycon Mobile gas analyzers and telemetry Temperature has significant impact on oxygen consumption. For example, in [13] authors tested with eleven healthy subjects at various temperatures. They found that, on average, oxygen consumption was 20.5% higher at −10 • C compared to 20 • C at same exercise intensities. Similarly, Sandsund et al. [14] discovered that at −15 • C oxygen consumption was 10.8% higher than at 23 • C during submaximal exercise intensities. They tested with eight elite non-asthmatic cross-country skiers. Regarding the above mentioned percentages it is crucial to note that the impact of temperature decreases as exercise intensity increases due to the body's increased heat production [15]. Oksa et al. [16] argue that the increase in oxygen consumption at lower temperatures may be caused by "shivering or increase thermoregulatory tonus of the muscles", altered neuromuscular function and use of additional clothing, which increases the metabolic rate.
Influence of humidity on oxygen uptake was investigated in [17]. The authors of this study found no difference inVO 2 at four sub-maximal velocities of 2.7, 3.3, 3.9, and 4.5 m/s, respectively, for 4 min per stage during tests with different levels of relative humidity (23, 43, 52, 61 and 71%) at 31 • C temperature. As our measurements were taken in May at 20 • C/40% relative humidity (https://www.jyv-weather.info/wxhistory.php?date=202205, (accessed on 13 January 2023)) and in November at 5 • C/90% relative humidity, the absolute humidity, i.e., the total mass of water vapor present in the air were almost the same. Thus, in our measurements humidity has no considerable impact on theVO 2 measurements. VO 2 measurements and respiratory frequency were collected using breath-by-breath methods from a portable spirometer (Oxycon Mobile, Jaeger, Würzburg, Germany) every 5 s. Immediately before every test session, the gas analyzer and the flow meter were calibrated. The data from the Oxycon Mobile were collected on the notebook using provided telemetry. The accuracy of Oxycon Mobile was discussed in [18]. According to the authors of this study, the Oxycon Mobile significantly underestimatesVO 2 at high workloads above 200 W. Typical bias is about 100 mL/min and standard deviation is approximately 125 mL/min. The oxygen uptake measurements were post-processed to reduce the measurement noise and remove outliers. A Savitzky-Golay filter [19] with polynomial order and frame length set to 3 and 1, respectively, was applied three times on the oxygen input data (Figure 2). HR measurements were recorded continuously (beat-by-beat) during the test with a Suunto Movesense IMU+HR system with an output rate of approximately 1 Hz. It was also smoothed and interpolated after the tests.
Measurements of running and walking parameters were collected continuously by the INS/GPS datalogger and saved to a memory card through a wired connection, which ensured that no data were lost. For our outdoor walking and running tests on a level track, the datalogger unit was attached to the torso of test subjects. After the experiment, the data were transmitted to cloud storage using a 4G/LTE USB modem connected to the datalogger. A detailed description of this setup is provided in [3]. The accuracy of speed and speed difference is approximately 0.05 m/s. Vertical displacement and step duration are computed with the accuracy of about 0.01 m and 10 ms, respectively. The output frequency for acceleration, velocity, angular velocity and orientation is 400 Hz. To compute walking and running metrics, the step segmentation was performed. The metrics were computed with the step frequency for each step.
Dataset Preparation
Oxygen uptake, heart rate and motion measurements were synchronized in time in a post-processing phase. Jumps with both feet were executed at the beginning of each data recording simultaneously with bookmarks in oxygen uptake measurements, to obtain time-synchronization between the three different devices. The jumps caused sharp peaks in the acceleration measured both by the INS/GPS and the Suunto Movesense IMU+HR systems. Using these peaks, the synchronization of oxygen uptake, heart rate and motion measurements, as well as the data analysis, were performed offline on a computer after the experiment. The achieved accuracy in time synchronization of about 0.1-0.2 s is sufficient because HR andVO 2 are slowly changing variables with sampling rates of 1 Hz and 0.2 Hz, respectively.
The motion data were processed to obtain step segmentation and compute running and walking metrics for each step as described in [3]. Accelerations and velocities were computed in the anatomical frame. Once the data were segmented into steps, features or metrics that are commonly used in walking and running were computed for each step to facilitate analysis of the data (see [3,11] for details). The oxygen uptake and HR measurements were resampled to meet the same step-by-step frequency as the running metrics. To use an LSTM neural network for oxygen uptake prediction, a training set of sequences (input features) and target values (oxygen uptake) must be created. The following running metrics were computed for each step and selected as input features in this study: • Speed averaged over one step: arithmetic mean of the speed (=step length/duration of step), m/s; • Speed: peak-to-peak difference during one step, m/s; • Step duration, s; • Vertical displacement: peak-to-peak difference in vertical movement, m; • Heart rate, bpm.
The number of steps in input sequences is one of the hyperparameters that relates the length of past surrogate variables and oxygen uptake. Based on experiments, we decided that input sequences of 50 steps yield a good estimate of the time-dependence decay between the output and past values of inputs. For a better fit, and to prevent the training from diverging, the input sequences were normalized between 0 and 1.
In machine learning, optimal feature selection is crucial for developing simple yet reliable prediction models. The purpose of feature engineering is to identify valid, useful, and understandable patterns in INS/GPS data that have a strong correlation with the target parameter (VO 2 ), but minimum inter-correlation with the other features. Figure 3 shows the absolute correlations between pairs of input features and target features averaged for all participants. These recommendations are not binding. Sometimes features that are weakly correlated with the target features can significantly improve the machine learning algorithm prediction accuracy if they are combined with other features. Furthermore, one needs to keep in mind that weak correlation between the input and the target features indicates only that there is no linear relationship between the two features but provides no information about potential non-linear relations.
The selected set of input features was validated based on the consider only one and leave-one-out approaches. It shows that speed, speed change and heart rate are the most important input vectors to predict theVO 2 label. Adding vertical displacement and step duration improves the predictions to some extent. Adding further input features does not improve the prediction accuracies. Figure 3 shows that speed and oxygen uptake are correlated. However, if speed or HR is the only input feature the accuracy is worse than when they are combined with other features. The optimal set of input features was selected based on correlation matrix and our domain knowledge. The entire dataset was shuffled and split into training and test subsets with 80% to 20% ratio.
LSTM Network Architecture
Oxygen uptake predictions have to take into account the historical context of inputs, therefore a many-to-one long-short term memory (LSTM) model was developed for oxygen uptake (VO 2 ) prediction. Optimal length of the input sequences is estimated based on the results. We tried sequences of 50 and 100 steps and concluded that shorter sequences are better. The neural network was implemented in Matlab Deep Learning toolbox. The neural network includes the following layers: one LSTM layer, input, output and dense layers. We used a sequence input layer that can contain two to five input sequences and its input size matches the number of channels of the input data. The LSTM layer has 150 hidden units. The number of hidden units determines how much information is learned by the layer. Larger values can yield more accurate results but can be more susceptible to overfitting to the training data. The output is a single number for each set of sequences giving the value for predictedVO 2 . A fully connected layer with a size matching the number of predictors, followed by a regression layer, were used to specify the number of values to predict. The number of LSTM layers is a hyperparameter and was selected based on experiments. The total number of trainable parameters was 93,151 that is far more than training samples (typically 1500-5000). However, LSTM networks work very well despite these potential overfitting problems, mainly because of various regularization effects implicit to the training/optimization algorithm. The obvious benefit of having many parameters is that the network is flexible enough to represent the desired mapping. Adapting the size of the network to the size of the training set can lead to a problem when the network is too simple and unable to represent the desired mapping (high bias).
It was trained using the Adam optimizer for 8000 epochs. For larger datasets, you might not need to train for as many epochs for a good fit. The learning rate was set to 0.005. Training the neural network required approximately 3-5 h on a PC equipped with an Intel Core i7-6700 @3.4 GHz CPU processor. Testing the models requires only a few seconds for every simulation.
To assess the prediction ability of the different models, a residual analysis was conducted. Residuals were calculated as the difference between the experimentalVO 2 values and the outputVO 2 values predicted by the models. The RMSE of the residuals was calculated. A Bland-Altman analysis was used to assess the level of agreement between measured and predicted data. The mean bias and the limits of agreement at 95% of probability (two times standard deviation) were calculated.
Input Features Include Only Motion Parameters
Two different input feature options were compared here: (a) four motion parameters including speed, speed change (peak-to-peak) during one step, step duration and vertical oscillation; (b) only speed. The Bland-Altman analysis of the predictedVO 2 using LSTM network with four motion input features and directly measuredVO 2 across all exercise conditions and participants combined are shown collectively in Figure 4. Dashed horizontal lines represent 95% limits of agreement and the solid line represents the prediction bias. Each color represents data from a unique participant in the test set. This plot shows that the prediction bias is −0.04 mL/min/kg (approximately 0.2% ofVO 2peak ) and the validity of the predictedVO 2 values from the LSTM network expressed with 95% limits of agreement is 2.49 mL/min/kg (approximately 5% ofVO 2peak ).
Example ofVO 2 prediction accuracy for a single representative athlete is shown in Figure 5. The plot showsVO 2 measurements (blue) and predictions (red). Based on our tests with 11 participants, it can be concluded that the bias is typically small, less than 0.5 mL/min/kg (<1% ofVO 2peak ). The standard deviation (σ) is typically about 1.2-1.5 mL/min/kg, 95% of all predictions are within 2σ. The Bland-Altman analysis of the predictedVO 2 using LSTM network with one input feature (speed) and directly measuredVO 2 across all exercise conditions and participants combined are shown collectively in Figure 6. Performance of the same LSTM network for a single representative athlete is shown in Figure 7. The LSTM network with four input features yields more accurate prediction than the LSTM network with only one feature: 2.49 vs. 5.79 mL/min/kg (95% limits of agreement). Although the vertical displacement, speed change and step duration cannot be used alone, they significantly improve the prediction accuracy if they are combined with speed ( Figure 5).
Input Features Include Only Heart Rate
When heart rate is the only input feature to the LSTM network, the Bland-Altman analysis of the predicted and directly measured oxygen uptake across all exercise conditions and participants combined is shown collectively in Figure 8. This plot shows that the prediction bias is 0.16 mL/min/kg (approximately 0.35% ofVO 2peak ) and the validity of the predictedVO 2 values from the LSTM network expressed with 95% limits of agreement is 2.52 mL/min/kg (approximately 5.5% ofVO 2peak ). Performance of the regressor for a single representative athlete during 20 min-long test is shown in Figure 9.
Input Features Include Motion Parameters and Heart Rate
When the input features to the LSTM network include both the motion features and heart rate, the Bland-Altman analysis of the predicted and directly measured oxygen uptake across all exercise conditions and participants combined is shown collectively in Figure 10. This plot shows that the prediction bias is 0.02 mL/min/kg (approximately 0.05% ofVO 2peak ) and the validity of the predictedVO 2 values from the LSTM network expressed with 95% limits of agreement is 1.36 mL/min/kg (approximately 3% ofVO 2peak ). Performance of the regressor for a single representative athlete during 20 min-long test is shown in Figure 11.
Discussion
Our hypothesis was that an LSTM neural network could be used to accurately predict individual's oxygen uptake during walking and running from measurements of motion parameters and heart rate. We also wanted to examine how heart rate measurements could improveVO 2 prediction accuracy if they are added to the wearable INS/GPS device that measures acceleration, velocity, angular velocity and orientation of the upper body where the device is attached.
To the best of our knowledge, we are the first to apply recurrent neural networks for prediction ofVO 2 during unconstrained walking and running based on measurements of motion parameters only and motion parameters combined with the heart rate. The results show that performance of the LSTM network with the input features that include only motion parameters and the LSTM network with heart rate as the only input are comparable: the bias is −0.04 mL/min/kg vs. −0.16 mL/min/kg and the 95% limits of agreement is 2.49 mL/min/kg vs. 2.52 mL/min/kg. In the first case the bias is negligible. If the motion features are combined with heart rate the LSTM network's performance is considerably better: the bias is 0.02 mL/min/kg and the 95% limits of agreement is 1.36 mL/min/kg.
All previous studies used heart rate and sometimes mechanical power (in cycling ergometer case) for oxygen uptake prediction. In such approaches, mechanical power is a very important feature. However, it cannot be measured for walking and running directly as it is measured on cycling ergometer. For example, Zignoli et al. [4] studied the ability of recurrent neural networks to predict oxygen uptake during exercises on cycling ergometer. They show that, using heart rate, mechanical power output, pedaling cadence and respiratory frequency as input features to the LSTM network, it is possible to estimateVO 2 with the accuracy of 7.2% ofVO 2max which is approximately 3.5 mL/min/kg. The performance of our approach forVO 2 prediction during unconstrained walking and running is better (1.47 mL/min/kg). Some studies did not use the mechanical power, but the prediction was less accurate. For example, Amelard et al. [5] developed temporal convolutional neural network to predictVO 2 during cycle ergometer exercise using only cardiorespiratory bio-signals (HR, HR reserve, breathing frequency, and minute ventilation). The prediction accuracy was 8% ofVO 2max which is approximately 3.8 mL/min/kg. As our LSTM model's performance was mainly evaluated using data collected during the same session, it remains unclear how well it can perform with data collected during different days. There are many factors that can influence daily oxygen uptake during exercise, such as previous physical activity, state of health and contents and the time of previous meal. Even when these things are taken into account, day-to-day variation in oxygen consumption can be around 5% when performing exercise under a constant load [20]. When thinking about changes in oxygen consumption during exercise in the long run, factors such as endurance training, resistance training and nutrition can also have an effect [21,22].
In our approach, the model training is based on data from a single person and prediction is performed for the same person, similar to other studies [4,5]. Despite of this limitation, our approach still has practical importance. Once the training is complete, the system can be used for several months until a brief update is required. During this time, without using expensive and cumbersome equipment (portable spirometers), it can estimate fitness level, physical conditions, and shape (including potential injuries). To develop a robust model that is stable with respect to physiological variability, the training data have to include measurements from several tests carried out at different temperatures and times over a long period of time. The development of generalizable models for inter-subject (i.e., using trials of some subjects in model training, and taking the trials of different subjects for validation) oxygen uptake prediction is still an open research question.
Conclusions
Our study suggested that an LSTM network can predict oxygen consumption based only on sequences of motion data collected from wearable sensors with an accuracy of 2.49 mL/min/kg (95% limits of agreement). Motion input features have to include speed and speed change. The addition of vertical oscillations and step duration does not provide a substantial improvement. Physiological markers, such as heart rate, can be added to the input and improve theVO 2 prediction accuracy (1.36 mL/min/kg 95% limits of agreement). Achievable accuracy is comparable with the measurement accuracy of an Oxycon Mobile (approximately 250 mL/min 95% limits of agreement) portable spirometer when the training and predictions are for the same person and performed at similar temperatures. Based on this study and our experience withVO 2 prediction, we think that this algorithm has the potential to be embedded in a portable system and to provide real-time assessment of individualVO 2 during walking and running. The proposed approach forVO 2 prediction can provide a unique opportunity for continuedVO 2 collections in unsupervised environments. Our approach can be applied in continuous assessment of energy expenditure and aerobic fitness with the potential for future applications such as the early detection of possible injuries and the deterioration of physical health. The proposed algorithm can also be adapted to estimate energy expenditure and the quantification of training intensity. We investigated exercising conditions at low and moderate intensities. More work is needed to cover all different conditions including heavy and severe intensity exercises. The effect of inter-subject variation and day-to-day variability onVO 2 prediction has yet to be studied. | 6,829.2 | 2023-02-01T00:00:00.000 | [
"Engineering",
"Computer Science",
"Medicine"
] |
Recognition of Mixture Control Chart Pattern Using Multiclass Support Vector Machine and Genetic Algorithm Based on Statistical and Shape Features
Control charts have been widely utilized for monitoring process variation in numerous applications. Abnormal patterns exhibited by control charts imply certain potentially assignable causes that may deteriorate the process performance. Most of the previous studies are concerned with the recognition of single abnormal control chart patterns (CCPs). This paper introduces an intelligent hybrid model for recognizing the mixture CCPs that includes three main aspects: feature extraction, classifier, and parameters optimization. In the feature extraction, statistical and shape features of observation data are used in the data input to get the effective data for the classifier. A multiclass support vector machine (MSVM) applies for recognizing the mixture CCPs. Finally, genetic algorithm (GA) is utilized to optimize the MSVM classifier by searching the best values of the parameters of MSVM and kernel function.The performance of the hybrid approach is evaluated by simulation experiments, and simulation results demonstrate that the proposed approach is able to effectively recognize mixture CCPs.
Introduction
In today's manufacturing and service industries, control charts are particularly important tools to improve product quality and monitor production process.Various kinds of control charts have been developed by different quality attributes and control targets.Recognizing control chart patterns (CCPs) is one of the most prevalently used techniques to detect process disturbances, equipment malfunctions, or other special events.In general, six basic CCPs are commonly exhibited by control charts, including normal (NOR), cyclic (CC), increasing trend (IT), decreasing trend (DT), upward shift (US), and downward shift (DS).Figure 1 shows these six types of control chart patterns [1].Over the past two decades, attention has been given to improve the recognition accuracies of these basic CCPs using normalized original data.Automatic CCPs recognition was an active research area in last decade but has not yet been realized fully.
There are numerous research papers on CCPs organization.Most of the previous studies are concerned with the recognition of single abnormal CCPs [2][3][4].However, in practice, the observed process data may be mixture CCPs, which may be combined with two or three basic patterns.Compared to the basic patterns, the mixture patterns are more difficult to recognize and result in serious performance degradation for patterns recognition.So it is a challenging task to identify mixture patterns effectively.Only a few studies have reported on mixture patterns recognition [5][6][7][8].Guh and Tannock [5] use the back-propagation neural network to recognize the mixture CCPs.H. Yang and S. Yang [6] propose an efficient statistical correlation coefficient method for the recognition of mixture CCPs.Chen et al. [7] integrate wavelet method and back-propagation neural network for online recognition of mixture CCPs.Lu et al. [8] propose a hybrid system that uses independent component analysis and supports vector machine to recognition mixture CCPs.
Feature extraction plays an important role in CCPs recognition.Most of the existing literatures use normalized original data as the inputs.These data normally generate large structures and are not very effective for complicated recognition problems.A smaller data size can lead to faster And other feature extraction methods are proposed for eliminating the duplicated information, like independent component analysis (ICA) [11], fisher discriminate analysis (FDA) [12], and principal component analysis (PCA) [13,14].The feature extraction efforts cited above did not approach a suitable set of features.In this paper, thirteen features that consist of both statistical and shape features of the CCPs are initially chosen.It is a well-established dimensionality reduction technique, which can be employed to compress the noise and correlated measurements, so that makes the data into a simpler and smaller informative subspace for measurement data sets.Traditionally, CCPs were analyzed and interpreted manually.Until the end of the 1980s, expert systems were employed for control chart patterns recognition [15,16].With the development of computer technology, machine learning techniques have been widely adopted in automatic process monitoring.In particular, artificial neural networks (ANNs) are the most frequently used in control chart patterns recognition [17][18][19][20].The use of ANNs has overcome some drawbacks in the traditional expert system method.Artificial neural networks utilize a multilayer perception with back propagation training to classify unnatural patterns and show higher accuracy.In subsequent studies, many other methods like decision tree, fuzzy clustering, and wavelet analysis are combined with ANNs to recognize CCPs [19,20].
However, ANNs also suffer from several weaknesses, such as the need for a large amount of training data, bad generalization ability, the risk of model over-fitting, difficulty to obtain stable solution, and getting into a local extremum easily.The application of ANNs is limited due to these weaknesses.Support vector machine (SVM), based on statistical learning theory, is proposed to recognize CCPs because of its excellent performance in the practical application.It mainly used the principle of structural risk minimization, which makes it have greater generalization ability when there is a small sample, and is superior to the principle of the empirical risk minimization principle as artificial neural networks [7,21,22].The biggest problems encountered in setting up the SVM model are how to select the kernel function and its parameters values.The parameter set of the penalty parameter and kernel function parameter should be optimized.
The purpose of this study is to develop an intelligent hybrid CCPs recognition model that can be used for mixture CCPs to improve the recognition accuracy.This paper considers the six basic and four mixture CCPs and generates their statistical and shape features as the inputs and multiclass support vector machine (MSVM) as classifier.At the same time, genetic algorithm (GA) is chosen as an optimization tool to optimize the MSVM parameters.This model will improve CCPs recognition performance.
Modeling for Control Chart Patterns Recognition
The aim of this model is to recognize CCPs effectively and automatically.Figure 2 shows the schematic diagram representing the procedure of the CCPs recognition, in which three modules are in series: feature extraction, classifier, and parameters optimization (F MSVM GA).
In the feature extraction module, statistical and shape features of observation data are used as the data inputs for the classifier.As we know, every control chart pattern has different properties, and features represent the properties of various CCPs.If some effective features are chosen to reflect the pattern, it is easier to recognize the abnormal patterns.Original data as the inputs usually have large data and are not very effective for complicated recognition problems.In this paper, statistical and shape features of CCPs as the feature extraction method are utilized to get the suitable data.In classifier module, an MSVM classifier is developed for recognizing the basic and mixture patterns.In order to achieve satisfactory recognition performance, the MSVM classifier needs to be properly designed, trained, and tested.However, using MSVM has some difficulties, like how to select the optimal kernel function type and the most appropriate hyperparameters values for MSVM training and testing stages.Therefore, genetic algorithm is applied for finding the optimum values of hyperparameters, that is, the kernel parameter and classifier parameters in parameters optimization module.
Statistical and Shape
Features.The patterns can be described in the original data.The statistical features and shape features can be got from the original data.It is efficient to simplify the data number and get the useful information.In this paper, eight statistical features and five shape features are chosen to reflect the patterns; these thirteen features are, respectively, shown below [2].
(1) Mean.The mean for normal and cyclic pattern is around zero, while that for other patterns is different from zero.
Therefore, it may be a good candidate to differentiate normal and cyclic patterns from other patterns: (2) Standard Deviation.Standard deviation of sample data, each mode performance is different (4) Average Autocorrelation.This paper takes the average of correlation degree between property values for each sample: (5) Positive Cusum.Sample data points are greater than the average and then cumulate the gap between data and their average: (6) Negative Cusum.Sample data points are smaller than the average and then cumulate the gap between data and their average: (7) Skewness.It provides information regarding the degree of asymmetry: (8) Kurtosis.It measures the relative peakness or flatness of its distribution: 2.1.2.Shape Features.They are as follows: slope, N1, N2, APML, and APLS.
(1) Slope.The slope of the least-square line: the slope for normal and cyclic pattern is around zero, while that for other patterns is greater than zero.Therefore, it may be a good candidate to differentiate normal and cyclic patterns from other patterns: (2) N1.The number of mean crossing: it is almost zero for shift and trend patterns but very high for normal patterns; cyclic pattern is the intermediate pattern.The feature differences can distinguish the normal and cyclic from shift and trend patterns: (3) N2.The number of least-square line crossing: this feature is the highest for normal and trend patterns, intermediate for shift patterns, and the lowest for cyclic patterns.Thus it can be used for separation of normal and trend patterns from others: (4) APML.The area between the pattern and its mean line: this feature is the lowest for normal pattern; therefore, it differentiates the normal pattern from others: (5) APLS.The area between the pattern and its least-square line: normal and trend patterns have lower values than shift and cyclic patterns.Thus it can be used to distinguish normal and trend patterns from shift and cyclic patterns:
Support Vector Machine. Basic SVM is invented by
Vapnik of the AT&T Bell lab team.It is created based on the VC dimension theory and structural risk minimization of statistical learning theory.So that gets the best solution between model complexities and learning ability according to the limited sample information.The basic SVM deals with two-class problems.However, it can be extended to Multiclass SVM [7,[21][22][23]].
An SVM performs classification tasks by constructing optimal separating hyperplanes (OSHs).An OSH maximizes the margin between the two nearest data points belonging to two separate classes.Suppose that the training set, ( , ), = 1, 2 . . ., , ∈ , ∈ {−1,1}, can be separated by the hyperplane ⋅ + = 0, where is the number of sample observations and is the dimension of each observation and is the weight vector and is the bias.If this hyperplane separates the data from two classes with maximal margin width 2/‖‖ 2 and all the points on the boundary are named the support vector, the SVM solves the following optimization problem: This is a convex quadratic programming (QP) problem, and Lagrange multipliers ( , = 1, 2, . . ., ; ≥ 0) are used to solve it.And, for input data with a high noise level, an SVM using soft margins can be expressed with the introduction of the nonnegative slack variables ( , = 1, 2 . . ., ).Equation ( 14) is transformed into the following constrained form: In (15), is the penalty factor; it determines the penalty degree of the error.It can be viewed as a tuning parameter, which can be utilized to control the trade-off between maximizing the margin and the classification error.
An MSVM method is adopted in the classifier stage.There are two methods: one-against-all (OAA) or one-against-one (OAO).Suppose that it has an N-class pattern recognition problem; N independent SVMs are constructed and each of them is trained to separate one class of samples from all others.When testing the system after all the SVMs are trained, a sample is input to all the SVMs.Suppose that this sample belongs to class N1; ideally only the SVM trained to separate class N1 from the others can have a positive response.Another method is called one-against-one (OAO) method.For an N-class problem, SVMs are constructed and each of them is trained to separate one class from another class.Again, the decision of a testing sample is based on the voting results of these SVMs.In this paper, OAO is adopted for patterns recognition [24].
In the nonlinearly separable cases, which cannot be linearly separated in the input space, the SVM uses the kernel method to transform the original input space into a high dimensional feature space, where an optimal linear separating hyperplane can be found.Although there are several types of kernel function, the most widely used kernel function is the radial basis function (RBF), which is defined as The largest problems encountered in MSVM are how to select the penalty parameter () and kernel function parameters value ().The GA is used to search for the best value of parameters in MSVM classifier.
Genetic Algorithms (GA).
GA is a powerful tool in the field of global optimization.It has better search efficiency, robustness, and parallel compared with traditional optimization algorithms.Genetic algorithms belong to the larger class of evolutionary algorithms, which generate solutions to optimization problems using techniques inspired by natural evolution, such as inheritance, mutation, selection, and crossover.
In this paper, OAO and RBF kernel function are adopted for MSVM; the performance of an MSVM is mainly impacted by the setting of parameters of two parameters ( and ).The GA is used to search for the best value of parameters in MSVM classifier.The particle has two dimensions: and ; the accuracy of training set is selected as the fitness function.The steps are as follows [25].
Step 1. Set GA parameters, like the number of population, evolutionary generation, crossover and mutation probability, and parameter ranges.
Step 2. Optimize coding parameters and initialize the population.
Step 3. Optimize decoding parameters and calculate the recognition rate, decode chromosomes of population, select the training set of model, and use training set recognition rate as the fitness function of the GA algorithm, so that we obtain the optimal MSVM parameters ( and ).
Step 4. Genetic manipulation (selection, crossover, and mutation): each chromosome does the selection, crossover, and mutation based on the fitness, thus excluding low fitness chromosomes and leaving high fitness chromosomes.The new group members are outstanding in the previous generation groups, which are better than the previous generation.GA performs iteratively until meeting some predetermined optimized targets.
Step 5. Get optimal parameters: decode the best chromosome, use the optimal parameters to train the training data in support vector machine classifier, and ultimately get the optimized support vector machine classifier.
Simulation and Results Analysis
3.1.Data Generation.In order to analyze the CCPs recognition, Monte Carlo method is used to get the sample data.The following equation is applied to generate the data points for six basic patterns; different parameters are shown in Table 1 [8]: where () means the value of sample data at time ; is the mean of data; () = × , is the random value of standard normal distributed between −1 and 1, is the standard deviation of normal distribution, and () is the abnormal value.We chose = 0, = 1 and use the 40 data points of observation window as inputs of the feature extraction model.Every pattern generates 100 sample data.However, the observed process data may be mixture control chart patterns in practice, which is combined with two or three basic patterns.Figure 3 shows four kinds of mixture CCPs, which are combined with cyclic, increasing trend and decreasing shift.We know that the principles of increasing/decreasing trend or upward/downward shift are similar, so the increasing trend and downward shift are chosen for the mixture CCPs.And sample data of mixture CCPs can be generated by different parameters in Table 3. Six basic CCPs (see Figure 1) and four mixture CCPs (see Figure 3) are, respectively, used for training and testing the proposed F MSVM GA method in this study (Table 2)., so as to cover high or small regulations of the classifier and fat or thin kernels, respectively.In the GA optimization module, there are several coefficients, whose values can be adjusted to produce better performances during training in this study, are summarized in Table 3.
Performance Analyses
In this section, we measure the performance of the proposed recognizer.For this purpose, we have previously generated 10 patterns, 100 of each type; every sample has 40 data points of observation windows.And we have used about 50% of the sample for training the classifier and the rest for testing.The testing samples can be used to estimate the performance of recognizer for each pattern and then compute the average recognition accuracy of CCPs.Several performances are done to verify the effectiveness of the proposed model.
Performance of Recognizer in Optimization.
First, we have applied MSVM classifier with different features.
Table 4 indicates the recognition accuracy (RA) of proposed F MSVM GA model on the 13 statistical and shape features and GA optimization algorithm.In order to demonstrate the superior performance of the proposed F MSVM GA scheme, MSVM using 13 features as inputs without GA optimization (called F MSVM) is constructed; the performance results are shown in Table 5.
As reported in Tables 4 and 5, the average recognition accuracies of F MSVM GA and F MSVM are 97.4% and 91.4%.The proposed F MSVM GA model has better recognition performance for the mixture CCPs, especially in TS and CTS.Genetic algorithm searches for the best combination of MSVM classifier parameters to gain the fitness maximum, so as to improve recognition rate of testing samples.
Performance of Recognizer in Different
Features.Feature extraction can lead to faster training and more efficiency in CCPs recognition.Thirteen statistical and shape features are utilized as the inputs in this paper.In order to explain its effectiveness, MSVM classifier using the original 40 data points as the inputs (called D MSVM) is constructed.Table 6 shows the recognition accuracy of mixture CCPs.
The average recognition accuracies of D MSVM (78.0%) and F MSVM (91.4%) show that feature extraction method plays an important role in improving the recognition accuracy.From the data, we can find that mixture control chart patterns are difficult to recognize due to the complex relation, but the result is much better after using statistical and shape features extraction method.multilayer perception with back-propagation training to classify abnormal patterns.Back-propagation (BP) method is a common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent.We define that the number of input neurons is 50 and the number of hidden layers is 5. Table 7 shows the performance results.Compare MSVM models; the accuracy of BP method is only 65.4%, much lower than the other MSVM methods.The reason is that BP neural network quite depends on the quantity and quality of the sample data, but only 50 training samples are considered in this study; it belongs to the small sample noise problem.
We have compared the proposed model with other approaches.This comparison can be seen in Figure 4, and six basic CCPs and mixture CCPs are, respectively, numbered from 1 to 10.
Conclusion
Control charts are the most useful tools in statistical process control, and mixture control chart patterns are more and results indicate that the intelligent hybrid method can achieve the highest average recognition accuracies in the tested methods.
The future work will be focused on the following aspects: (1) employing statistical and shape features method as feature extraction method which we will compare with other excellent feature extraction methods, (2) comparing GA with other intelligent algorithms, including particle swarm optimization, simulated annealing algorithm, and ant colony optimization, (3) researching the fundamental principles of mixture CCPs with the help of mathematicians, and (4) seeking economic explanation of our method with the help of economists.
Figure 2 :
Figure 2: Flow chart of F MSVM GA model.
Table 3 :
Parameters in GA.
Table 5 :
Recognition accuracies with the F MSVM model.
Table 6 :
Recognition accuracies with the D MSVM model.
Table 7 :
Recognition accuracies with the BP model.widely used in manufacturing and service processes.Recognizing the mixture CCPs plays an important role in finding the abnormal quality problems.In this study, a hybrid method by integrating statistical and shape features extraction, MSVM, and GA are presented for recognizing the mixture CCPs.The proposed method initially uses statistical and shape features to get effective input data; then the combination of MSVM and GA is applied to recognize the mixture patterns.GA is to optimize the parameters of MSVM kernel parameters.Six basic CCPs and four mixture CCPs are used in this study for evaluating the performance of the proposed method.From the experiments, the simulation more | 4,625.6 | 2015-10-05T00:00:00.000 | [
"Computer Science",
"Engineering"
] |
Continuous Production of Biodiesel from Rubber Seed Oil Using a Packed Bed Reactor with BaCl 2 Impregnated CaO as Catalyst
The goal of this research was to test barium chloride (BaCl2) impregnated calcined razor clam shell as a solid catalyst for transesterification of rubber seed oil (RSO) in a packed bed reactor (PBR). The waste razor clam shells were crushed, ground, and calcined at 900 °C in a furnace for 2 h to derive calcium oxide (CaO) particles. Subsequently, the calcined shells were impregnated with BaCl2 by wet impregnation method and recalcined at 300 °C for 2 h. The synthesized catalyst was characterized by X-ray diffraction (XRD), scanning electron microscope (SEM), energy dispersive spectrometer (EDS), Brunauer-Emmett-Teller (BET) surface area, and basic strength measurements. The effects of various parameters such as residence time, reaction temperature, methanol/oil molar ratio, and catalyst bed length on the yield of fatty acid methyl ester (FAME) were determined. The BaCl2/CaO catalyst exhibited much higher catalytic activity and stability than CaO catalyst influenced by the basicity of the doped catalyst. The maximum fatty acid methyl ester yield was 98.7 % under optimum conditions (residence time 2.0 h, reaction temperature 60 °C, methanol/oil molar ratio 12:1, and catalyst bed length 200 mm). After 6 consecutive reactions without any treatment, fatty acid methyl ester yield reduced to 83.1 %. The option of using waste razor clam shell for the production of transesterification catalysts could have economic benefits to the aquaculture and food industries. Copyright © 2018 BCREC Group. All rights reserved.
Introduction
The increasing energy demand, depleting fossil fuels and environmental problem resulted from climate change by continues consumption of petroleum derived fuel, make challenges for the scientific community and the researchers worldwide today [1].Among the options explored for alternative energy sources, biodiesel is one of the attractive alternatives due to its renewability, biodegradability, and non-toxicity [2].It is usually produced via esterification or transesterification using acidic or basic catalysts.The most common process is to facilitate a reaction between triglycerides (TGs) and alcohol in the presence of a homogeneous catalyst [3].However, existing biodiesel processes suffer from some serious problems with the use of homogeneous catalysts, such as equipment corrosion, waste effluent treatment, soap formation and catalyst removal, leading to severe economic and environmental penalties.Therefore, exploring heterogeneous catalysts is becoming more important in chemical and life science industry [4].
Among the heterogeneous catalysts, calcium oxide (CaO) is a potential candidate for its low cost, high basic strength and catalytic activity, low methanol solubility, and green material.Calcium nitrate (Ca(NO3)2), calcium carbonate (CaCO3), β-tricalcium phosphate (β-Ca3(PO4)2), and calcium hydroxide (Ca(OH)2) are raw materials to produce CaO.Generally, CaO could be derived from CaCO3, which is the major constituent in many natural sources and wastes [5].Catalysts derived from waste shells were proposed for biodiesel production and it was concluded that these catalysts were environment friendly as they are mainly prepared from natural materials [6].Moreover, different kinds of CaO catalyst, such as gel-combusted CaO catalyst, CaO extracted from biowaste and supported CaO catalyst, had been investigated aiming at improving the catalytic activity [7].
Doping of metal oxides increases basic strength.Barium ions (Ba 2+ ) being highly basic, when added to the calcium ions (Ca 2+ ), become a competitive base catalyst with high conversions of TGs to biodiesel at lower reaction conditions [6,8].Based on these findings, we hypothesize that impregnating calcined razor clam shell surfaces with barium chloride (BaCl2) will enhance the catalytic activity of waste shells towards transesterification due to synergistic effects of a basic element deposited on a naturally basic solid support.Studies in which calcined razor clam shells have also been reported but to the best of our knowledge investigations have not been carried out to enhance their catalytic activity by impregnating it with Ba 2+ .
Generally, there are four major types of feedstock available for the biodiesel production including oil seed (vegetable oil), animal fats, algae and different low-quality material such as waste cooking oil (WCO), greases and soap stock [9].Non-edible plant oils have been found to be promising crude oils for the production of biodiesel.The use of non-edible oils when com-pared with edible oils is very significant in developing countries because of the tremendous demand for edible oils as food, and they are far too expensive to be used as fuel at present [10].Rubber seeds are the important by-product of the rubber trees and treated as a waste.These seeds contain 40-50 % oil and still underutilized in Thailand.Thus, oil extracted from the rubber seeds can be used as a prominent feedstock for the production of biodiesel in developing countries, where almost 85 % of the crude oil imports from the other countries [11].In the recent years, there are several works reported on biodiesel production using rubber seed oil (RSO) which has higher potential to be used as alternative diesel fuel because it is a non-edible oil that can produce sufficient amount of oil for the industry [12].Hence, in the present study, RSO is used as a feedstock for the synthesis of biodiesel.
Batch reactors are widely employed in the early research of biodiesel production.Compared with batch reactors, continuous reactors can reduce production cost, provide a uniform quality of the final product and are more conducive to large-scale industrial production.Among all reactors, packed bed reactors (PBR) are the most commonly employed reactors for biodiesel industrial production [13].There are lots of studies focused on using a PBR to improve the yields of biodiesel from vegetable oil.However, the RSO-derived biodiesel production using a PBR and novel catalyst BaCl2/CaO derived from waste razor clam shell has seldom been addressed.The present research aims to exploit BaCl2 impregnated CaO as a packed bed catalyst for the continuous production of biodiesel from RSO.The effects of residence time, reaction temperature, methanol/oil molar ratio, and catalyst bed length are systematically studied.Moreover, the reusability of the catalysts derived from the natural calcium materials was investigated.
Materials
The razor clam shell was collected as waste from Tae-lea Thai market, Phetchaburi Province, Thailand.The waste shell was rinsed with running water (H2O) to eliminate dust and impurities, and was then dried in an oven at 60 °C for 24 h.BaCl2 (AR grade) and all necessary solvents were procured from Ajax Finechem Pty Limited, Australia and were used as received without further purification.The rubber seed was collected from the Northeastern provinces of Thailand as an oil source for biodiesel production.The RSO was extracted using solvent extraction technique.It extracted with hexane is a clear yellow liquid.The molecular weight and density of the yellow oil were measured to be 875 g/mole and 0.887 g/cm 3 , respectively.
Preparation of catalyst
The waste razor clam shell was repeatedly washed to remove any organic impurities attached to it and then dried in an oven.The dried shell was crushed and sieved to pass 60-200 mesh screens (75-250 mm).The waste shell was calcined at a high temperature of 900 °C for 2 h in an air atmosphere with a heating rate of 10 °C/min similar to the procedure described by Buasri et al. [5].The calcined sample was obtained as white powder.Subsequently the CaO was impregnated with 2 M of BaCl2 as described by Mahesh et al. [8].BaCl2/CaO catalyst was prepared by the wet impregnation method, and was then dried in the vacuum oven at 60 °C for 24 h and recalcined at 300 °C for 2 h.The product was kept in the closed vessel to avoid the reaction with carbon dioxide (CO2) and humidity in the air before used.
Characterization of catalyst
X-ray diffraction (XRD) patterns of the samples were recorded on a LabX XRD-6100 analyzer (Shimadzu, Japan) operating at 30 kV and 20 mA with a Cu anode and a graphite monochromator (l = 1.5405Å), an angle of scanning range 10-70° (2θ), a scan step size of 0.04° and a scan rate of 4°/min.The XRD phases were identified using the Powder Diffraction File (PDF) database created by International Centre for Diffraction Data (ICDD).The sample morphology and elemental chemical analysis were also characterized at room temperature by a Hitachi TM3030 (USA) scanning electron microscopy (SEM) system equipped with an energy dispersive spectroscopy (EDS) detector.The material was coated with gold (Au) using a Sputter Coater for protecting the induction of the electric current.The accelerating voltage was 15 kV and working distance was 14 -15 mm.The surface area and pore distribution of the synthesized catalysts were determined by the Brunauer-Emmett-Teller (BET) method using a nitrogen (N2) adsorption/desorption analyzer (Autosorb-1 Model No. ASIMP.VP4, USA).Prior to the analysis, the catalysts were degassed at 300 °C for 4 h to remove moisture and foreign gases on the surface.Adsorption and desorption process of N2 on the catalyst surfaces were examined in a vacuum chamber at -196 °C.The basic strength of the synthesized catalysts was characterized using a Hammett indicators method as described by Boro et al. [6].
Continuous production of biodiesel
Continuous transesterification was performed in a PBR at an atmospheric pressure similar to the procedure described by Buasri and Loryuenyong [14] and Gui et al. [4].The reactor was composed of a water-jacketed stainless steel column with an external diameter of 60 mm, an internal diameter of 40 mm, and a length of 345 mm.The column was packed with BaCl2/CaO catalyst.The scheme of the continuous production of biodiesel from RSO and methanol is shown in Figure 1.RSO and methanol were charged into the system using a plunger pump.The reactants were mixed and preheated in a mixing column with random packing.The reaction temperature was controlled by a heater to keep a constant temperature of the reactor wall with an error of ±1.0 °C.The temperature difference between the inlet and the outlet was below 1.0 °C during all of the runs.Temperature and pressure of the system were monitored by a temperature indicator and pressure gauges.The effects of residence time (0.5 to 2.5 h), reaction temperature (50 to 70 °C), methanol/oil molar ratio (6:1 to 18:1), catalyst bed length (100 to 300 mm), and repeated operation of PBR (1-8 cycles) on the % yield of fatty acid methyl ester (FAME) were investigated.All experiments were repeated 3 times and the standard deviation was never higher than 7 % for any point.
The content of FAME in biodiesel samples was analyzed by the gas chromatographymass spectrometry (GC-MS, QP2010 Plus, Shimadzu Corporation, Japan) equipped with a flame ionization detector and a capillary GC column (DB-WAX, Carbowax 20M, 30 m × 0.32 mm × 0.25 μm) using inner standard method as described by Buasri et al. [15].The yield of biodiesel was calculated by dividing the weight of biodiesel with the weight of oil and multiplying the resulting number by 100 (Equation (1)). (1) The physical and chemical properties of FAME were analyzed according to ASTM and EN methods such as kinematic viscosity, density, flash point, cloud point, pour point, acid value, moisture content, ester content, free glycerine, total glycerine, iodine number,
Catalyst characterization
XRD patterns of natural razor clam shell, calcined shell and synthesized catalyst are given in Figure 2. The waste shell mainly consisted of CaCO3 phase, and after calcined at 900 °C for 2 h, this CaCO3 was completely converted to CaO by evolving the CO2.The formation of the sharp peak indicates the formation of highly crystalline materials [16].The catalyst shows the presence of CaO, Ca(OH)2, and BaCl2 with prominent 2θ values at 18.08°, 29.36°, 34.06° and 47.12° is also seen.Doping with barium (Ba) brings some shifting in the XRD pattern of the parent CaO.The nonvisibility of the peaks in the doped samples might have arisen due to the fact that the d value of diffraction peaks of tetragonal phase barium oxide (BaO) is very close to the cubic phase CaO.It might be that the diffraction peaks of both the phases overlapped almost completely [6].The XRD analyzes confirmed the presence of BaCl2 on the shell surface.
The morphology of natural razor clam shell, calcined shell and the synthesized catalyst was detected by SEM (4000x magnification) and results were showed in Figure 3. From the figure, we can see that the morphology of natural shell displayed a typical layered architecture with macro-pores and irregular shapes of particles of various sizes.When the calcined temperature was 900 °C, the particle shapes become more regular.The SEM images of CaO showed a cluster of well-developed cubic crystal with obvious edges [17].It was observed that the particles of the catalyst have a spherical structure and are agglomerated into lumps.This is because the sample is in the oxide form.The isolated particles were reasonably isotropic [2,8].
Elemental analysis of the synthesized catalyst was characterized by EDS which is well known analytical technique to determine the chemical or elemental characterization of a material [18].Figure 4 represents the EDS of BaCl2/CaO derived from waste razor clam shell.
Here, the positions of specific elements emitting characteristic X-rays within an inspection field can be indicated by unique color.The weight percent (norm.wt.%) of calcium (Ca), oxygen (O), barium (Ba), and chlorine (Cl) on a particular area at the surface of catalyst were determined to be 78.08 %, 18.04 %, 1.16 % and 2.72 %, respectively.The weight ratio between barium and calcium (Ba:Ca) was 1.49, while the weight ratio between chlorine and calcium (Cl:Ca) was 3.48, indicating that BaCl2 was well loaded onto CaO.Furthermore, the surface of the catalyst has strong basic character and therefore holds tendency to absorb moisture.H2O adsorbed within the pores on the basic catalyst surface has an influence on O:Ca.Hydrogen was not quantified by EDS because of small orbital diameter (atomic number = 1), which reduces catch probability for spectroscopy to almost zero [5].
The BET specific surface area (S BET ), average pore diameter (D p ) and total pore volume (V p ) of the calcined shell and synthesized catalyst are summarized in Table 1.It indicated that the CaO catalyst (calcined shell) possessed a surface area of 18.78 m 2 /g, a pore diameter of 9.77 nm and a pore volume of 0.053 cm 3 /g.The results of BaCl2/CaO catalyst showed gradually reduced of surface area and pore volume upon increasing the Ba species concentration.The significant reduction in the S BET and V p for synthesized catalyst was due to the excess of active metals embedded into the channel of parent materials and incorporated into the pore of CaO [1].The average pore size of the novel catalyst was 1.16 time bigger than that of the CaO catalyst, suggesting that the particles of the calcined and treated razor clam shell were packed much closer.
Hammett indicator method was used to determine the basic strength of the catalyst according to the literature [6].The basicity and basic strength are the most important properties regarding their application as a synthesized catalyst for the transesterification of RSO.The reaction activity depends on the number of basic sites present in the catalysts as well as on their strength [5].The increasing Hammett indicator (15.0 < H_ < 18.4) and basic strength (0.238 mmol/g) for doped catalyst might have resulted due to the presence of BaO and CaO active species (Table 2).
Effects of transesterification process variables
The catalytic activities of the packed bed BaCl2/CaO catalysts were studied by investigating the effects of residence time, reaction temperature, methanol/oil molar ratio, catalyst bed length, and reusability of catalyst in the transesterification.For the following experiments, the calcined and treated razor clam shell was used as a novel catalyst to catalyze the reaction of RSO and methanol in PBR.
The effect of residence time on the conversion of RSO to FAME was investigated.Figure 5 shows a gradual increase in the yield of FAME with time from 0.5 h to 2.5 h with a catalyst bed length of 200 mm and a methanol/ oil molar ratio of 12:1.The FAME yield of 98.7 % was obtained in 2.0 h at 60 °C for BaCl2/CaO catalyst.Production of biodiesel increases with residence time until it reaches equilibrium at a residence time of 2.0 h, due to the initial transient of the reactor from the start-up to the stationary conditions [14].It has also been estimated that for the operation beyond 2.0 h, the yield increases slightly.This is because longer residence time results in hydrolysis of esters (side reaction) and a reversible reaction between RSO and methanol [19].In this study, the optimum residence time appears to be 2.0 h.
In order to determine the optimum reaction temperature, transesterification was carried out by varying the reaction temperature from 50 to 70 °C in a step increment of 5 °C (Figure 6).The low conversion was obtained when the reaction was carried out at low temperature of 50-55 °C which is attributed to the mass transfer and solubility limitations.At higher reaction temperature (60 °C) it is expected that the reactants may gain sufficient kinetic energy that will accelerate the mass transfer rate among the RSO-methanol catalyst phases resulting in maximum conversion.However, the slight loss in activity is observed when the reaction temperature is above 60 °C.This is probably due to the fact that methanol has a boiling point near to 65 °C and reaction above this temperature might cause the loss of methanol leading to low catalytic activity [6].In addition, the catalytic activity of BaCl2/CaO decreases due to some qualitative changes, such as loss of Ba 2+ and Ca 2+ .Moreover, in order to save energy, it is necessary to choose the relative low temperature [20].Therefore, the optimum reaction temperature for the transesterification of RSO to FAME is considered to be around 60 °C.
The transesterification with methanol/oil molar ratio of 6:1 to 12:1 (Figure 7) indicates that the conversion of RSO to FAME increases rapidly with the mole ratio, reaching 98.7 % at the mole ratio of 12:1 catalyzed by BaCl2/CaO.Further increasing mole ratio to 18:1, the conversion increases gradually to 99.2 %.In addition, stable emulsions may form at the higher content of methanol (18:1), leading to complicated separation and purification of biodiesel [4].At the end of transesterification, the excess methanol was recovered by a vacuum distillatory and recycled in latter reactions.
The catalyst bed length in the PBR is associated with the residence time during continuous production of biodiesel.The RSO conversion to FAME increased rapidly from 65.4 to 81.2 % with the increase of the catalyst bed length from 100 to 150 mm (Figure 8).In particular, the yield of FAME was over 98 % when the catalyst bed length was over 150 mm.This demonstrates that a high catalyst bed length provides a lower flow rate and more active sites to promote the reaction between RSO and methanol at a given residence time [14,21].All results presented hereafter were measured using a 250 mm bed length.
The divalent metal oxide catalysts having a substantial amount of covalent character facilitates the transesterification reaction [22] as depicted in Figure 9.In general, the catalytic properties are influenced by particle size, surface area, as well as the active sites of a catalyst.Small sized catalyst with high surface area and active sites will give rise to high catalytic performance [23].In this study, Ba has been used to increase the basicity of the CaO derived from the waste razor clam shell.The increase in the yield of FAME with Ba loading is attributed to the presence CaO and BaO which might have acted like active species during the reaction [6].
Catalyst reusability
One of the main disadvantages of homogeneous catalysts is that they cannot be recovered.So, in the present study, the catalyst was evaluated to study its efficiency and reusability.The catalyst was recycled to test its activity as well as stability.Reusability of the BaCl2/CaO catalyst in the transesterification of RSO was checked without any further purification and activation.The solid catalyst was collected after adding fresh reactants.The catalytic activity remained 83.1 % of the fresh catalyst when the novel catalyst was employed for the 6 consecutive runs (Figure 10).It was calculated based on the conversion of RSO to FAME under the optimum condition (residence time 2.0 h; reaction temperature 60 °C; methanol/oil molar ratio of 12:1; and catalyst bed length 200 mm) in the PBR.The decay in cata- lyst activity could be due to the leaching of active sites to the reaction media.Notwithstanding the interaction between the Ba species and CaO surface, the loaded species are leached by the reaction media.Leaching of the active phase to the alcoholic phase can be attributed to the bond breaking and formation of Ca 2+ and CH3O - [24,25].
Fuel properties of biodiesel
For biodiesel to be used in diesel engines, the fuel must meet various specifications stated in biodiesel standard, mainly United States biodiesel standard (ASTM D-6751) and European biodiesel standard (EN 14214) [5].The fuel properties (kinematic viscosity, density, flash point, cloud point, pour point, acid value, moisture content, ester content, free glycerine, total glycerine, iodine number, cetane number and calorific value) of FAME obtained in this work are summarized in Table 3 along with a comparison to the recommended biodiesel international standards ASTM D-6751 and EN 14214.It can be seen that most of its properties are in the range of fuel properties as described in the latest standards for biodiesel.
Conclusions
The present study demonstrates the successful application of calcined and treated waste razor clam shell as the efficient heterogeneous catalysts for transesterification of RSO with methanol in PBR.The highest FAME yield of 98.7 % for BaCl2/CaO catalyst was obtained under the optimum condition (residence time 2.0 h; reaction temperature 60 °C; methanol/oil molar ratio of 12:1; and catalyst bed length 200 mm).The reusability studies of novel catalyst showed continuous drop in catalytic activity which was attributed to the loss of active sites and deposition of product on the catalyst itself.The RSO conversions maintained higher than 80 % in 6 consecutive runs.The experimental results showed that novel catalysts had excellent catalytic activity, stability, and reusability during the reaction.Moreover, this new continuous biodiesel production process and apparatus have good potential for utilization of RSO with an inexpensive and easily available non-edible oil.
Fuel Property
Biodiesel number, and calorific value of the biodiesel.The obtained values were then compared with the United States biodiesel standard (ASTM D-6751) and European biodiesel standard (EN 14214).
Figure 1 .
Figure 1.Scheme of the continuous production of biodiesel from RSO and methanol with a PBR
Figure 3 .
Figure 3. SEM micrographs of (a) waste razor clam shell (b) calcined shell and (c) novel catalyst.Images are magnified at 4000X, and scale bars represent 20 mm (c) (b)
Figure 4 .
Figure 4. EDS analysis of novel catalyst: (a) chemical mapping for Ca, O, Ba, Cl, and (b) weight percent (norm.wt.%) of elements on surface
Figure 5 .Figure 6 .
Figure 5.Effect of residence time on the %yield of FAME.Reaction conditions: reaction temperature 60 °C; methanol/oil molar ratio 12:1; and catalyst bed length 200 mm
Figure 9 .
Figure 9. Mechanism of a novel base heterogeneous catalyst during transesterification
Figure 7 .Figure 8 .
Figure 7. Effect of methanol/oil molar ratio on the %yield of FAME.Reaction conditions: residence time 2.0 h; reaction temperature 60 °C; and catalyst bed length 200 mm
Table 2 .
Hammett indicator and basic strength of calcined shell and novel catalyst
Table 1 .
S BET , D p , and V p of calcined shell and novel catalyst
Table 3 .
The fuel properties of biodiesel obtained in the transesterification of RSO Figure 10.Effect of reusability of catalyst on the %yield of FAME.Reaction conditions: residence time 2.0 h; reaction temperature 60 °C; methanol/oil molar ratio of 12:1; and catalyst bed length 200 mm | 5,430.6 | 2018-06-11T00:00:00.000 | [
"Chemistry"
] |
Different Effects of Guanine Nucleotides (GDP and GTP) on Protein-Mediated Mitochondrial Proton Leak
In this study, we compared the influence of GDP and GTP on isolated mitochondria respiring under conditions favoring oxidative phosphorylation (OXPHOS) and under conditions excluding this process, i.e., in the presence of carboxyatractyloside, an adenine nucleotide translocase inhibitor, and/or oligomycin, an FOF1-ATP synthase inhibitor. Using mitochondria isolated from rat kidney and human endothelial cells, we found that the action of GDP and GTP can differ diametrically depending on the conditions. Namely, under conditions favoring OXPHOS, both in the absence and presence of linoleic acid, an activator of uncoupling proteins (UCPs), the addition of 1 mM GDP resulted in the state 4 (non-phosphorylating respiration)-state 3 (phosphorylating respiration) transition, which is characteristic of ADP oxidative phosphorylation. In contrast, the addition of 1 mM GTP resulted in a decrease in the respiratory rate and an increase in the membrane potential, which is characteristic of UCP inhibition. The stimulatory effect of GDP, but not GTP, was also observed in inside-out submitochondrial particles prepared from rat kidney mitochondria. However, the effects of GDP and GTP were more similar in the presence of OXPHOS inhibitors. The importance of these observations in connection with the action of UCPs, adenine nucleotide translocase (or other carboxyatractyloside-sensitive carriers), carboxyatractyloside- and purine nucleotide-insensitive carriers, as well as nucleoside-diphosphate kinase (NDPK) are considered. Because the measurements favoring oxidative phosphorylation better reflect in vivo conditions, our study strongly supports the idea that GDP cannot be considered a significant physiological inhibitor of UCP. Moreover, it appears that, under native conditions, GTP functions as a more efficient UCP inhibitor than GDP and ATP.
Introduction
The mitochondrial proton electrochemical gradient generated by the respiratory chain pumps drives ATP synthesis as a result of oxidative phosphorylation (OXPHOS). However, F O F 1 -ATP synthase is not the only factor consuming this gradient. The inner mitochondrial membrane (IMM) possesses many protein carriers included in the mitochondrial anion carrier protein family, among which uncoupling proteins (UCPs) and adenine nucleotide translocase (ANT) both have an affinity to bind purine nucleotides as well as to mediate non-phosphorylating proton leak [1,2]. UCPs have been identified across eukaryotes, including eukaryotic microorganisms, plants, and vertebrate as well as invertebrate species [3]. UCPs specialize in proton translocation from the intermembrane space into the mitochondrial matrix, a process that is not related to ATP synthesis and dissipates energy as heat. The abundant expression and action of UCP1, the first described UCP, in brown adipose tissue of small, hibernating, and coldacclimated mammals are responsible for the thermogenic properties of this tissue. However, the role of UCP isoforms in unicellular organisms, such as the amoeba Acanthamoeba castellanii, and in the non-thermogenic cells of plants and animals remains unclear [3]. As UCPs, similar to artificial uncouplers, decrease the mitochondrial membrane potential (DY) and thus the redox state of respiratory chain electron carriers, their action should lower the production of reactive oxygen species [4]. Although many data indicate the antioxidative action of UCPs, including UCP1 [5][6][7][8], some results do not confirm the involvement of UCPs as oxidative stress response proteins [9]. What is interesting, recently it has been proposed that UCP2 can function as a metabolite (malate, oxaloacetate, and aspartate) transporter [10]. The UCP2-mediated metabolite transport is coupled with the exchange for phosphate plus a proton. Thus, taking into account earlier reports focused on UCP2 function [4,7], proton transport through the UCP2 could be coupled with the metabolite transport or not. Other carriers of IMM, e.g. ANT, also mediate metabolite transport and mitochondrial uncoupling [2]. In turn, ANTdependent adenine nucleotide turnover (the exchange of intramitochondrial ATP for extramitochondrial ADP) is necessary to maintain OXPHOS in mitochondria. In addition to the basic function of ANT, this carrier can also mediate proton leak and behave like a classic uncoupling protein [2]. UCP-mediated proton leak and ANT-mediated proton leak need to be finely controlled to prevent ATP depletion in the cell. UCPs, as well as ANT, are stimulated by free fatty acids [1,11] and hydroxynonenal, a membrane lipid peroxidation end product [12], though the inhibition of futile protein-mediated (ANT-or UCP-mediated) proton conductance is crucial for efficient ATP synthesis in mitochondria. Carboxyatractyloside (CATR), a specific inhibitor of ANT, inhibits purine nucleotide transport and ANT-mediated proton leak [11]; interestingly, ANT even fully inhibited by CATR can still contribute to the proton conductance [13]. Purine nucleotides (PNs) were first recognized as specific inhibitors of UCPs [1], and PN-dependent inhibition of respiratory rate accompanied by the restoration of DY is considered diagnostic of UCP function in isolated mitochondria. However, unlike the ANT-mediated proton conductance, which can be inhibited by ADP and GDP [11,14,15], UCPs are strongly inhibited by ATP and GTP, in addition to ADP and GDP. Therefore, because of this difference in specificity, the inhibition of mitochondrial proton conductance by PNs other than ADP and GDP should be considered diagnostic of UCP function [16]. ANT specializes in ADP and ATP transport across IMM, but nucleotide analogs may bind to the ANT without being transported [2]. For example, GDP is considered a weak binding competitor for ANT. Although GDP and GTP are known to accumulate in the mitochondrial matrix, the specific and ANT-independent mechanism of guanine nucleotide import into mitochondria is poorly described [17][18][19].
Interpretations of the effect of GDP on isolated mitochondria often omit at least two very important phenomena, namely: (i) the presence of nucleoside-diphosphate kinase (NDPK) in mitochondria, e.g., in the intermembrane space [20] and (ii) the possibility of GDP oxidative phosphorylation in mitochondria [21,22]. NDPK is an enzyme that catalyzes the transfer of a c-phosphate group from ATP (and other nucleoside triphosphates) to nucleoside diphosphates, e.g., ATP + GDP R ADP + GTP [23]. In mammals, only the NDPK-D isoform possesses a mitochondrial targeting sequence and is ubiquitously expressed, with the highest expression in liver, kidney, bladder, and prostate [20]. Thus, in light of the common usage of GDP in studies of UCP-mediated and ANT-mediated uncoupling, mitochondrial GDP turnover, i.e., GDP accumulation in the mitochondrial matrix [17][18][19], GDP transphosphorylation [23], or GDP oxidative phosphorylation [21,22], complicate the interpretation of the GDP inhibitory effect on protein-mediated mitochondrial uncoupling.
Most functional studies on the activity of UCPs have been performed in non-phosphorylating (state 4) respiration, i.e., in the presence of oligomycin (an inhibitor of F O F 1 -ATP synthase), which prevents mitochondrial OXPHOS. However, these assays with isolated mitochondria do not reflect physiological conditions in which phosphorylating (state 3) respiration and non-phosphorylating (state 4) respiration are mixed. The main objective of the present work was to elucidate the action of GDP and GTP in isolated mammalian mitochondria under physiological-like conditions, i.e., those favoring OXPHOS (without the inhibitors CATR and oligomycin) and in the presence of ATP, which is abundantly synthesized in mitochondria. We suggest that under such conditions, the GDP-dependent inhibition of UCP should be severely weakened because of the action of NDPK and/or because of direct GDP import into the mitochondrial matrix, enabling its oxidative phosphorylation.
Animals
The experiments were carried out on adult 8-10-week-old male Wistar rats weighting 250-350 g. The animals were bred in the animal house at the Poznan University of Medical Sciences, Poznan, Poland. They were given free access to water and pellet food and were housed under standard humidity and temperature conditions on a 12
Isolation of mitochondria
For each experiment, kidney mitochondria were isolated from three rats. Isolation procedure was performed on ice in the cold room. The kidneys were washed, comminuted, and homogenized by four passes with a glass-Teflon homogenizer in ice-cold isolation medium containing 100 mM sucrose, 100 mM KCl, 50 mM Tris-HCl, 1 mM KH 2 PO 4 , 0.1 mM EGTA, 0.5 mM EDTA, and 0.2% fatty acid-free bovine serum albumin (BSA), pH 7.2. The presence of BSA in the isolation medium allowed the endogenous free fatty acids to be chelated from the homogenate suspension. The homogenate was filtered through sterile gauze, and the filtrate was centrifuged at 700 x g for 10 min at 4uC. The supernatant was centrifuged at 10 000 x g for 10 min at 4uC. Lowspeed centrifugation at 700 x g for 10 min at 4uC preceded a final high-speed centrifugation at 8 000 x g for 10 min at 4uC. The final mitochondrial pellet was resuspended in ice-cold storage buffer containing 225 mM mannitol, 75 mM sucrose, 0.1 mM EDTA, and 10 mM Tris-HCl, pH 7.2. The mitochondrial protein concentration was determined by the biuret method with BSA as a standard.
Preparation of inside-out submitochondrial particles (SMP) from rat kidney mitochondria The inside-out SMP were prepared according to published procedures with some modifications [25,26]. The pellet of rat kidney mitochondria was resuspended in ice-cold high-salt medium (225 mM mannitol, 75 mM sucrose, 10 mM Tris-HCl, and 20 mM MgCl 2 , pH 7.2.) to a final volume of approximately 4 ml with 10 mg protein/ml. Mitochondria sonication was done in a glass beaker (with a flat bottom) placed in an ice bath with a Bandelin Sonopuls sonifier for four 10 s bursts (50% of power) separated by a 1 min cooling period. After mitochondria disruption, the sample was diluted to approximately 20 ml with the medium used for sonication and centrifuged at 8 000 x g for 10 min at 4uC. The relatively intact mitochondria from the pellet were resuspended in a high-salt medium and further disrupted by a repetition of the sonication and pelleting procedure described above. The resulting two supernatants were then centrifuged at 20 000 x g for 10 min at 4uC to pellet intact mitochondria and large membrane fragments. Finally, supernatants were centrifuged at 105 000 g for 60 min at 4uC (Beckman Coulter Optima XPN-100 Ultracentrifuge, 70Ti rotor). The resulting pellets (SMP) were rinsed twice to reduce salt content, pooled and resuspended in icecold buffer containing 225 mM mannitol, 75 mM sucrose, and 10 mM Tris-HCl, pH 7.2. The SMP protein concentration was determined by the Bradford method with BSA as a standard.
Oxygen uptake was measured polarographically using a Clarktype oxygen electrode (Rank Brothers, Cambridge, UK) in 2. 1.5 mM EGTA, and 0.1% BSA. Succinate (5 mM) was used as an oxidizable substrate in the presence of rotenone (4 mM) to block electron input from complex I. As a control of mitochondrial quality, phosphorylating respiration was measured for each mitochondrial preparation to evaluate the coupling parameters.
Only high-quality mitochondrial preparations, i.e., those with an ADP/O value of approximately 1.3 (with succinate as a respiratory substrate) and a respiratory control ratio of approximately 3.6-3.9, were used in the experiments. The values of O 2 uptake are given in nanomoles of O per minute per milligram of protein. The mitochondrial membrane electrical potential (DY) was measured simultaneously with oxygen uptake using a tetraphenylphosphonium (TPP + )-specific electrode, as previously described [16]. The values of DY are given in millivolts (mV).
Measurement of SMP respiration
Oxygen uptake was measured polarographically with a Clarktype oxygen electrode (Hansatech Instruments, UK) at 37uC in 0.6 ml of incubation medium (225 mM mannitol, 75 mM sucrose, 10 mM Tris-HCl, 5 mM KH 2 PO 4 , 0.18 mM MgCl 2 , 0.5 mM EDTA, and 0.1% BSA, pH 7.2) with 0.4 mg of SMP protein. The exogenous NADH oxidation (having no place in intact mammalian mitochondria), which was almost completely sensitive to rotenone, as well as the insensitivity of ADP-stimulated and GDPstimulated respiration to carboxyatractyloside confirmed that our SMP were indeed inside-out SMP.
Phosphorylating respiration measurements
The ADP/O ratio was determined by an ADP pulse method using 400 nmol of ADP for intact mitochondria. The total amount of oxygen consumed during phosphorylating respiration was used to calculate the ratio. The simultaneous measurements of DY enabled the fine control of the duration of state 3. It was necessary to incubate isolated mitochondria with ADP or ATP to observe the OXPHOS-like effect after GDP addition (Figs. 1A, 1B, 1D and Fig. S1A in File S1); in the case of ADP, GDP was added in the post-ADP state 4. The maximal GDP-induced OXPHOS-like effect in intact mitochondria required the equivalent concentration of ADP or ATP and was observed at a low GDP concentration, i.e., 120 mM ( Fig. 1). On the contrary, the GDP-dependent stimulation of oxygen uptake in SMP did not require the presence of ATP (Fig. 2). However, the ADP-stimulated respiration and the GDP-stimulated respiration in SMP was much weaker compared to intact rat kidney mitochondria. It was not surprising, because it is generally agreed that SMP can be less coupled than intact mitochondria based on facts that usually in SMP: (i) ADP/O ratios are very low, (ii) respiratory control (state 3-state 4 transition) is barely observed and (iii) the ATPase activity is very high [27]. What is more, the same authors claim that the high rate of the backflow of protons via ATPase which is oligomycin-insensitive (''intrinsic uncoupled'' activity of ATPase) significantly affects the apparent coupling in OXPHOS experiments with SMP.
Proton leak measurements
The response of proton conductance to its driving force can be expressed as the relationship between the oxygen consumption rate and DY (a flux-force relationship) when varying the membrane potential by titration with respiratory chain inhibitors. The proton leak kinetics ( Fig. 3 and Fig. S1 in File S1) were examined in the presence of 0.8-1 mM ATP under four different conditions: (i) in the absence of CATR and oligomycin, (ii) in the presence of CATR (3.6 mM) alone, (iii) in the presence of oligomycin (0.7 mg/ml) alone, and (iv) with the simultaneous presence of CATR and oligomycin. To induce UCP activity, linoleic aid (LA) was used at a concentration of 25 mM (with 2 mg of mitochondrial protein in 2.8 ml of incubation medium). LA was always added on stabilized state 4 respiration (in the absence or presence of OXPHOS inhibitors). GDP and GTP were added to 1 mM, and always after LA. The respiratory rate and DY were varied by modulating the coenzyme Q-reducing pathway with malonate (0.3-1.6 mM), a competitive inhibitor of succinate dehydrogenase. To assess the statistical significance of the induced shifts in the proton leak curves, we generally compared the respiration rates at the highest common DY values for pairs of curves from 5-7 independent experiments using Student's t-test for unpaired data.
Statistical analysis
The results are presented as the means 6 S.D. obtained from at least 5 mitochondrial isolations, with each determination performed at least in duplicate. An unpaired two-tailed Student's ttest was used to identify significant differences; in particular, differences were considered to be statistically significant if p,0.05 (*), p,0.01 (**), or p,0.001 (***).
Results and Discussion
A high concentration of GDP (1 mM) stimulates the respiratory rate and decreases the membrane potential of isolated mitochondria The addition of GDP and other nucleoside diphosphates can stimulate respiration in isolated mitochondria when measurements are performed in the absence of OXPHOS inhibitors (oligomycin and/or CATR) [23,28]. This phenomenon can be explained as a result of NDPK activity, which can be bound to IMM [29]. Thus, in the presence of GDP and ATP, there is a possibility for local ADP regeneration as a consequence of the transphosphorylation reaction catalyzed by NDPK and the subsequent stimulation of OXPHOS [23]. Using mitochondria isolated from rat kidneys and human endothelial cells, we also observed the stimulation of the respiratory rate accompanied by a decrease in DY in the absence of OXPHOS inhibitors, but this effect occurred with a much higher GDP concentration, than previously described. The phenomenon can be observed in traces of simultaneous measurements of respiratory rate and DY (Figs. 1A, 1B and 1D) and fluxforce relationships, (Fig. 3A and Fig. S1A in File S1). Specifically, in the presence of 1 mM ATP, even a concentration of 1 mM GDP could induce a state 4-state 3 transition that was sensitive to CATR and oligomycin (Fig. 1A). However, as with 1 mM ADP, the state 3-state 4 transition, revealing the cessation of OXPHOS, was not observed with 1 mM GDP (data not shown).
Previous studies using isolated mammalian mitochondria have shown that, in contrast to other nucleoside diphosphates, GDP acts both as a substrate (up to 0.15 mM) and an inhibitor (above 0.15 mM) of NDPK, with almost complete inhibition by 0.6 mM GDP [23,30,31]. In our experiments, using 120 mM GDP (not inhibitory for NDPK), the stimulation of respiration (state 4-state 3 transition), followed by the return to basal respiration (state 3-state 4 transition), was clearly observed in the human endothelial mitochondria and rat kidney mitochondria (Figs. 1B and 1D). As shown for the rat kidney mitochondria, 1 mM GDP was still stimulating, though its effect revealed a weaker acceleration of the respiratory rate accompanied by a much smaller decrease in DY (Fig. 1A) in comparison to the conditions with a lower (120 mM) GDP concentration (Fig. 1B). Similar results were obtained using the human endothelial cells mitochondria (data not shown). In contrast to GDP, 1 mM GTP inhibited the respiratory rate and increased DY in the absence of OXPHOS inhibitors, revealing an inhibition of mitochondrial uncoupling most likely mediated by UCP (Fig. 1A, dashed traces). It also means that in tested mitochondria, GTP is poorly susceptible to dephosphorylation and is not a substrate for NDPK.
Because the action of mitochondrial NDPK is most likely fully inhibited in the presence of 0.6 mM GDP [23,30], the observed 1 mM GDP-induced state 4-state 3 transition (Figs. 1A, 3A and Fig. S1 in File S1) could be a result of GDP import into the mitochondria [17][18][19], followed by the GDP oxidative phosphorylation [21,22] rather than NDPK-dependent GDP transphosphorylation. However, previous studies on ANT specificity toward PN exchange have excluded guanine nucleotide transport [17,32,33]. Moreover, it has been shown that the transport of ADP through ANT can be competitively inhibited by GDP [34]. Nevertheless, GDP is readily taken up and concentrated in the matrix of isolated mitochondria through at least two distinct mechanisms, i.e., atractyloside-sensitive and -insensitive pathways [17,19].
Similar to a previous report [14], in the absence of an ATP or ADP prepulse, we did not observe the state 4-state 3 transition upon the addition of GDP (data not shown). The matrix ATP pool, generated upon ADP prepulse or direct ATP addition, was crucial for the GDP-induced OXPHOS-like effect in the absence of oligomycin and CATR (Figs. 1A, 1B, 1D, 3A and Fig. S1A in File S1). Similar to ADP, extramitochondrial ATP is an effective ligand for the exchange of intramitochondrial ATP via ANT. The uptake of ATP can also be catalyzed by the mitochondrial ATP-Mg/P i transporter that is little sensitive to CATR [35]. Thus, extramitochondrial ATP can effectively generate the matrix pool of ATP, which in turn might be exchanged with exogenous GDP via carrier-mediated transport. Moreover, the maximal GDPinduced OXPHOS-like effect for a given GDP concentration was strictly dependent on the ATP (or ADP pulse) concentration and required a 1:1 ratio (Figs. 1A, 1B, and 1D). This observation supports the idea that the putative carrier-mediated uptake of GDP could be coupled to the efflux of a counter-substrate and resembles ADP transport through ANT in exchange for ATP. Additionally, it must be kept in mind that OXPHOS of GDP is most likely conserved in all organisms. Bacterial F O F 1 -ATPase [36] and mammalian mitochondrial F O F 1 -ATP synthase [21,22] have an affinity to bind GDP and phosphorylate it to produce GTP.
A smaller decrease in DY and a weaker increase in the respiratory rate caused by GDP compared to ADP (Figs. 1B and 1D) could reflect the much lower affinity of ANT as well as F O F 1 -ATP synthase for GDP compared to ADP [2,21,22]. The failure to observe GDP-induced stimulation of the respiratory rate accompanied by a decrease in DY in the presence of CATR (Figs. 1C, 3C and Fig. S1C in File S1) might indicate the involvement of ANT in GDP translocation or, alternatively, an as-yet-unknown CATR-sensitive carrier. Irrespectively of the GDP concentration, the GDP effect was also sensitive to oligomycin (Figs. 1A, 1C, 3B and Fig. S1B in File S1) and required the presence of inorganic phosphate (data not shown), indicating the induction of OXPHOS upon GDP addition. It must be stressed that, under our experimental conditions, none of the applied GTP concentrations could induce the state 4-state 3 transition that is characteristic of the GDP OXPHOS-like effect in the absence of OXPHOS inhibitors (Figs. 1, 3A and Fig. S1A in File S1). Ideal conditions to test the potency of GDP to induce OXPHOS without the risk of its transphosphorylation should exclude the action of NDPK. Unfortunately, we could not switch off the activity of NDPK to determine the effect of GDP addition because one of the known inhibitors of NDPK, 39-azido-39deoxythymidine (AZT) [30], is also a strong inhibitor of ANT [37]. In turn, cromoglycate, also identified as an NDPK inhibitor, in as high a concentration as 10 mM can block NDPK catalytic activity by only 60% [38]. Therefore, the only way to completely preclude NDPK-dependent transphosphorylation without switching off ANT is to use a higher concentration of GDP [23,30]. Additionally, the usage of labeled GDP will not clarify where the newly synthesized GTP arose: GTP as a product of F O F 1 -ATP synthase activity resides in the mitochondrial matrix, where NDPK is also located [29]. The same could be true for GTP arising through NDPK activity bound to IMM but facing the mitochondrial intermembrane space, as GTP from this site can be transported to the matrix [19]. Thus, by performing measurements under conditions favoring OXPHOS, it is difficult to unambiguously determine whether the stimulation of respiration in isolated mitochondria upon GDP addition in the presence of ATP, particularly at lower GDP concentrations when NDPK is active, is merely the result of NDPK activity or due to GDP import into the mitochondria. The GDP-dependent stimulation of respiration (the state 4-state 3 transition) is most likely a result of a mixture of these two phenomena, a mixture of ADP (a product of NDPK activity) and GDP oxidative phosphorylation. Experiments with inside-out SMP (from rat kidney mitochondria) supported GDP OXPHOS phenomenon (Fig. 2). In SMP, ADP and GDP stimulated the respiratory rate, while GTP was ineffective. However, the potency of both nucleotides to stimulate respiration decreased significantly in SMP compared to intact mitochondria. Nonetheless, the ADP effect was twice as strong as GDP effect. The GDP effect was observed in the absence of ATP, excluding the involvement of NDPK activity. The addition of CATR did not inhibit the ADP-and GDP-induced acceleration of respiratory rate. Thus, the ADP and GDP effects were not governed by transporter systems (e.g., ANT). However, these effects were abolished by oligomycin, indicating the involvement of F O F 1 -ATP synthase. Although experiments with SMP indicated GDP OXPHOS in mitochondria, SMP model did not clarify the involvement of GDP/GTP transport in the observed phenomenon.
Different effects of GDP and GTP on LA-induced mitochondrial proton leak in the absence of OXPHOS inhibitors
UCP and ANT are considered as two main catalysts of futile proton leak in mammalian mitochondria [12][13][14]. The presence of UCP2 in isolated rat kidney and human endothelial cell mitochondria was confirmed by western blot analysis (Fig. S2 in File S1). However, the effect of GDP and GTP on LA-induced mitochondrial respiration under conditions favoring OXPHOS is poorly recognized. Employing four different incubation conditions, i.e., in the presence of ATP (i) with no OXPHOS inhibitors, (ii) with oligomycin alone, (iii) with CATR alone, and (iv) with both inhibitors together, we studied different effects of GDP and GTP on rat kidney mitochondrial proton leak by analyzing fluxforce relationships in the presence of LA, an activator of UCP (Fig. 3). The LA effect was comparable in the absence or presence of OXPHOS inhibitors. We used a 1 mM concentration of guanine nucleotides to provide optimal conditions for UCP inhibition and to attenuate the process of transphosphorylation catalyzed by NDPK [23,30]. Under conditions favoring OX-PHOS, in the absence of oligomycin and CATR, the effects of GDP and GTP on LA-induced proton leak differed diametrically (Fig. 3A). Namely, the effect of GDP was stimulatory, and the effect of GTP was inhibitory. Similar results were obtained in the absence of LA, i.e., for non-induced proton leak (Fig. S1A in File S1). Column plots (Fig. 3) present the values of the respiratory rates reflecting proton leak at the highest common DY, i.e., at 170.5 mV for the four studied conditions. In the presence of 25 mM LA, 1 mM GDP increased the respiratory rate by approximately 41%, and 1 mM GTP decreased the respiratory rate by approximately 22% (Fig. 3A). The stimulating effect of GDP, under conditions favoring OXPHOS, was observed in the absence or presence of linoleic acid (Figs. 1, 3A and Fig. S1A in File S1), thus OXPHOS system and NDPK activity seem to be unaffected, at least by used LA concentration (25 mM). The LAinduced proton leak, in the absence of OXPHOS inhibitors, was fully inhibited by GTP when the Q-reducing pathway was gradually decreased with malonate (Fig. 3A), as previously described [16,39,40], what is diagnostic for UCP action. Thus, it may be concluded that among potential carriers of IMM involved in uncoupling process [2], UCP is the main catalyst of LA-induced proton leak, at least in rat kidney mitochondria under physiological-like conditions.
In turn, under conditions preventing OXPHOS, the stimulatory effect of GDP on LA-induced proton leak was completely abolished and the inhibitory effect of GTP was seriously weakened (Figs. 3B, 3C, and 3D). Similar results were obtained in the absence of LA, i.e., for non-induced proton leak (Figs. S1B, S1C, and S1D in File S1). The attenuation of PNs inhibitory effect in the presence of OXPHOS inhibitors could mean that both inhibitors, oligomycin and CATR, most likely exhibit a nonspecific action on UCP and impair the PN-dependent inhibition of UCP but not UCP-dependent proton leak. These data support many results showing moderate or no GDP-dependent inhibitory effect of proton conductance in the presence of OXPHOS inhibitors [12,14,15,[41][42][43][44][45]. It is likely that CATR induces a conformational change in UCP, which partly or fully prevents the PN-dependent inhibition of UCP. An analogous explanation has been proposed for ANT [46]. Namely, the structure change of ANT under the influence of hydroxynonenal likely desensitizes the ANT-mediated proton leak to CATR. However, because in our conditions, LA was still stimulatory in the presence of OXPHOS inhibitors and this LA-induced uncoupling was poorly sensitive to guanine nucleotides, a LA-stimulated component (likely carrier), which is insensitive to GDP, GTP and CATR, could mediate the proton conductance, as previously proposed for hydroxynonenalinduced uncoupling [45]. On the contrary, the LA-induced GTP/ GDP-insensitive uncoupling in the presence of CATR could involve ANT, because ANT even fully inhibited by CATR can still contribute to the proton conductance [13]. If a CATR-insensitive domain of ANT stimulates proton leak in the presence of LA, and this domain is insensitive to guanine nucleotides (as observed in this study, Fig. 3C), it indicates that guanine nucleotides do not function as inhibitors of ANT-mediated uncoupling, as previously described [11,45]. On the other hand, efficient guanine nucleotide-dependent inhibition of proton leak in the presence of CATR and oligomycin has been found for many UCP homologues [12,16,24,40,41,47,48]. Alternatively, the suppression of PNs inhibitory effect, especially in the presence of oligomycin alone, could be explained as a result of guanine nucleotides binding to ANT, proposed for GDP [8,14,15,34], and/or their import into the mitochondrial matrix, proposed for GDP and GTP [17][18][19], what certainly lowers the PNs concentration in the mitochondrial intermembrane space where UCP faces its PN-binding site [1].
The specificity of GTP and GDP action in the absence or presence of LA under conditions favoring OXPHOS, as described in detail above for isolated rat kidney mitochondria, is also characteristic in mitochondria isolated from human endothelial cells (data not shown).
General discussion
The incubation of isolated mitochondria under non-phosphorylating conditions (in the presence of oligomycin and/or CATR) limits the interpretation of the physiological regulation of UCP isoforms and ANT, two major catalysts of the proton conductance in mitochondria [12][13][14]. In recent years, the non-additivity of GDP (the most widely used PN for UCP inhibition) and CATR (thus far considered to be highly specific only for ANT inhibition) has been reported for the inhibition of mitochondrial proton leak [12,15,45]. Moreover, the potential promiscuity of GDP and CATR influences the inhibitory effect of ANT-mediated and UCP-mediated proton leak [12,15,44,45,49]. Therefore, the indirect (or nonspecific) interaction of CATR with UCP was proposed [12,15,45]. Once UCP is inhibited by CATR, any additional inhibition by PN would not occur. A strongly limited guanine nucleotides effect in the presence of CATR ( Fig. 3C and Fig. S1C in File S1) could be explained as follows: (i) CATR, irrespectively of the fatty acid presence, nonspecifically binds with UCP and impairs the PN-dependent inhibition of UCP (desensitization of UCP to nucleotides by CATR) but not UCPdependent proton leak, (ii) IMM carriers other than UCP and ANT are stimulated by fatty acids and catalyze proton leak, which is insensitive to GDP, GTP and CATR; in this case, CATR could be considered not only as inhibitor of ANT-mediated uncoupling but also UCP-mediated uncoupling, (iii) fatty acid-induced GTP/ GDP-insensitive uncoupling in the presence of CATR is catalyzed both by CATR-insensitive ANT domain [13] and desensitized to nucleotides (by CATR) UCP, and finally, (iv) carrier-independent uncoupling mechanisms mediate the proton leak. If CATR inhibits UCP, the effect is rather nonspecific (or indirect), because bongkrekate (another specific inhibitor of ANT) also likely inhibits the UCP-dependent proton conductance [15]. The inhibitory effect of CATR towards other carrier (citrate carrier) than ANT has also been reported [50]. However, it has also been reported that CATR does not influence the GDP binding to UCP2 [12] and does not inhibit UCP1 action [34]. In summary, if CATR and GDP are not specific for their classical targets, measurements focused on UCP inhibition should not be performed in the presence of GDP and CATR.
By creating in vitro physiological-like conditions, i.e., excluding CATR and oligomycin from the incubation medium supplemented with ATP, we identified another reason not to use GDP for UCP inhibition. For the first time at the level of mitochondrial functional studies using mitochondria isolated from rat kidney and human endothelial cells, we demonstrate that a high concentration of GDP (1 mM) stimulated the respiratory rate and decreased DY in the presence of LA, a potent activator of UCP isoforms (Fig. 3A). We observed exactly the opposite effect than what was expected for UCP inhibition, whereby the recoupling effect was revealed as a respiratory rate decrease accompanied by the restoration of DY. The effect of 1 mM GDP was more similar to ADP OXPHOS in which oxygen consumption is increased and DY decreases (Figs. 1A, 1B and 1C). This is a rather new perspective of the action of GDP in mitochondria, taking into account its common usage as an inhibitor of UCP-mediated proton leak. The state 4state 3 transition with 1 mM GDP, in the absence of OXPHOS inhibitors, was also induced in the absence of LA (Fig. S1 in File S1). Moreover, under conditions favoring OXPHOS and in the presence of ATP, another guanine nucleotide commonly used for UCP inhibition, i.e. GTP, produced the expected effects diagnostic for UCP inhibition, regardless of the presence of LA (Figs. 1A, 3A and Fig. S1A in File S1).
The GDP-induced OXPHOS-like effect could be explained by NDPK-dependent transphosphorylation between GDP and ATP, which generates the ADP pool and subsequently induces OXPHOS (Fig. 4) [23,28,30]. However, another explanation might be needed, as NDPK was found to be sensitive to increasing concentrations of GDP, being almost completely inhibited at 0.6 mM GDP [23,30], and taking into account the stimulatory effect of 1 mM GDP (Figs. 1A, 3A and Fig. S1A in File S1). The alternative explanation, in accordance with earlier studies [17,19,21,22,36], involves the possibility of the direct transport of GDP across IMM and its OXPHOS in the mitochondrial matrix as the 1 mM GDP stimulatory effect was completely abolished in the presence of OXPHOS inhibitors (Figs. 1, 3, 4 and Fig. S1 in File S1). CATR sensitivity of the GDP-induced OXPHOS-like effect (Figs. 1C, 3C and Fig. S1C in File S1) indicates that putative carrier-dependent GDP transport across IMM might be consistent with ANT or another, as-yet-unknown, CATR-sensitive guanine nucleotide carrier. This idea is supported by the atractyloside-sensitive pathway of guanine nucleotide transport in mammalian mitochondria [19]. However, the ANT involvement in GDP import is controversial [17,32,51]. On the other hand, it must be mentioned that, although GDP is considered to be a binding but not transported ligand for ANT, the clear differentiation of binding from a slow uptake has rarely been tested [2].
Generally, little is known about how GDP enters the mitochondrial matrix. In the yeast Saccharomyces cerevisiae, a specific GTP/GDP carrier has been described, however poorly affected by the powerful inhibitors of ANT, CATR and bongkrekate [52]. It has also been reported that the human mitochondrial deoxynucleotide carrier, which is partly sensitive to CATR, could be involved in GDP transport [53].
The purpose of enhanced GTP synthesis in the mitochondrial matrix, via, e.g., the GDP OXPHOS, could be due to GTPdependent mitochondrial protein synthesis and the incorporation of GTP into various mitochondrial RNAs [54]. An important role of GTP in mitochondrial iron metabolism has also been considered [55]. GTP cannot arise through the conversion of inosine or adenine compounds or via uptake and synthesis from the salvage pathways using guanine or guanosine [18]. Thus, an additional source of GTP arising from GDP OXPHOS could complement the well-known pathway of GTP synthesis during the citric acid cycle and support the expression of the mitochondrial genome during intensive mitochondrial biogenesis in rapidly growing tissues. In the mitochondrial matrix, GTP may also arise through NDPK action [29], though this pathway consumes valuable ATP. In turn, GDP OXPHOS to GTP consumes inorganic phosphate, thereby favoring a high energy potential in the cell. Different pathways of GTP synthesis, such as the citric acid cycle, NDPK action, and GDP OXPHOS, could also play an important role during the fusion and fission of mitochondria and thus regulate GTP-dependent mitochondrial membrane dynamics [56]. The high concentration of GTP in mitochondria could also prevent UCP action and promote ATP synthesis via OXPHOS during periods of high ATP demand.
On the basis of our present study, the following main conclusions can be drawn. UCP, an IMM carrier specializing in futile proton conductance, and ANT or other CATR-sensitive PN carriers that mimic UCP action are differently regulated by PNs. A CATR-sensitive PN carrier (ANT or an as-yet-unknown carrier) and UCP could compete for GDP in the intermembrane space of mitochondria. However, in the case of UCP, the binding of GDP could lead to the inhibition of UCP action, whereas the CATRsensitive PN carrier could mediate the transport of GDP, its substrate, across IMM. It must be considered that ANT and most likely other putative PN carriers require a 1:1 exchange [2]. During many studies of proton leak kinetics, only one exogenous PN was typically used [8,14,15,34], which could be the reason of the putative ANT inhibition by GDP. If ANT is involved in GDP import, GDP inhibition of ANT-sustained proton leak might be an artifact resulting from measurements performed in the presence of oligomycin; F O F 1 -ATP synthase arrested by oligomycin could cause GDP accumulation in the mitochondrial matrix and thus limit the carrier-mediated GDP import from the intermembrane space. Moreover, when ANT permanently mediated adenine nucleotide transport, no or a very weak competitive effect of GDP toward the transported ADP was revealed [14,17,28,57]. Thus, it is difficult to consider GDP as a potent inhibitor of ANT, although the GDP inhibitory effect of ANT-mediated proton leak was observed [8,14,15,34]. However, it must be mentioned that GDP was also reported to be a completely ineffective inhibitor of ANTmediated uncoupling [11,12,45,48,58,59]. The same might be true for ADP, which has also been described as an inhibitor of ANT-mediated proton leak [11,48,58], because the inhibitory effect of ADP was revealed in the presence of oligomycin, with ADP being transported via ANT but without being phosphorylated in the matrix (no activity of F O F 1 -ATP synthase). Besides GDP and ADP, some inhibition of free fatty acid-induced ANTmediated uncoupling was found with long chain acyl-CoA [11]. However this kind of inhibition could have minor physiological significance [50]. Similar to other authors [11][12][13]58], in the presence of oligomycin, we also observed an inhibition of the respiratory rate accompanied by DY restoration upon CATR addition (data not shown), indicating the contribution of ANT to mitochondrial protein-mediated proton leak. However, CATR, a glycoside synthesized by some plants (e.g., Xanthium species) [60], is not a physiological regulator of animal ANT. In summary, the highly specific physiological inhibitor of ANT-mediated proton Extramitochondrial GDP can be transphosphorylated to GTP in the presence of ATP in a reaction catalyzed by mitochondrial NDPK. NDPK is bound to the cytosolic-facing or matrix-facing IMM leaflet but is shown separately for clarity. Alternatively, extramitochondrial GDP can be imported via ANT or another, as-yet-unknown, CATR-sensitive carrier into the mitochondrial matrix to enable its OXPHOS. In turn, extramitochondrial GTP, or GTP synthesized in the NDPK-dependent reaction, functions as the stronger UCP inhibitor (blunt-end solid arrow) than GDP (blunt-end dashed arrow). doi:10.1371/journal.pone.0098969.g004 leak or proton leak sustained through another CATR-sensitive carrier(s) is still not identified Finally, our results suggest that, under physiological conditions, GTP could play the role of the strongest UCP inhibitor, stronger than GDP and ATP. Thus, GTP rather than GDP should be used in bioenergetics studies in vitro as diagnostic inhibitor of UCP function.
Supporting Information
File S1 Figure S1, The effect of GDP (1 mM) and GTP (1 mM) on proton leak of rat kidney mitochondria in the absence of linoleic acid. The relationships between the respiratory rate and DY (proton leak kinetics) are shown. The measurements were performed in the absence of OXPHOS inhibitors (2I) (A), in the presence of oligomycin (+O) alone (B), in the presence of CATR (+ C) alone (C), and with the simultaneous presence of both inhibitors (+O +C) (D). The oxidation of succinate (5 mM) was gradually decreased by increasing the concentration of malonate (0.3-1.6 mM). The mitochondria (2 mg) were incubated in 2.8 ml of incubation medium supplemented with rotenone (4 mM) and ATP (0.8 mM). The inserts of column plots show the respiratory rates at the highest common DY (175 mV) for the same dataset. The values are the means 6 S.D. of 6 independent experiments (mitochondrial isolations). Figure S2, Western blot analysis of mitochondria isolated from rat kidney (RKM) and human endothelial cells (HEM) using anti-UCP2 (sc-6525, Santa Cruz Biotechnology) antibodies. Detection of mitochondrial marker (cytochrome oxidase, COXII, MS404, MitoScience) was performed as a control. Experimental conditions as in [24]. Different amounts of protein (50 or 100 mg) were loaded into each lane (as indicated). (PDF) | 9,059.4 | 2014-06-06T00:00:00.000 | [
"Biology",
"Chemistry",
"Computer Science"
] |
Photo-Realistic Image Dehazing and Verifying Networks via Complementary Adversarial Learning
Physical model-based dehazing methods cannot, in general, avoid environmental variables and undesired artifacts such as non-collected illuminance, halo and saturation since it is difficult to accurately estimate the amount of the illuminance, light transmission and airlight. Furthermore, the haze model estimation process requires very high computational complexity. To solve this problem by directly estimating the radiance of the haze images, we present a novel dehazing and verifying network (DVNet). In the dehazing procedure, we enhanced the clean images by using a correction network (CNet), which uses the ground truth to learn the haze network. Haze images are then restored through a haze network (HNet). Furthermore, a verifying method verifies the error of both CNet and HNet using a self-supervised learning method. Finally, the proposed complementary adversarial learning method can produce results more naturally. Note that the proposed discriminator and generators (HNet & CNet) can be learned via an unpaired dataset. Overall, the proposed DVNet can generate a better dehazed result than state-of-the-art approaches under various hazy conditions. Experimental results show that the DVNet outperforms state-of-the-art dehazing methods in most cases.
Introduction
In outdoor environments, acquired images lose important information such as contrast and salient edges because the particles attenuate the visible light. This degradation is referred to as hazy degradation, which distorts both spatial and color features and decreases visibility of the outdoor object. If the hazy degradation is not restored, we cannot expect a good performance of main image processing or image analysis methods such as object detection, image matching, and imaging systems [1][2][3][4], to name a few. Therefore, the common goal of dehazing algorithms is to enhance the edge and contrast while suppressing intensity or color saturation. To the best of the authors' knowledge, Middleton and Edgar were the first to employ a physical haze model for the dehazing problem [5].
To generate the haze-free image using the physical model, atmospheric light and the corresponding transmission should be estimated. However, an accurate estimation of the atmospheric light and transmission map generally requires additional information, such as a pair of polarized images, multiple images under different weather conditions, distance maps, or user interactions [6][7][8][9]. For that reason, many state-of-the-art approaches try to find a better method to estimate the atmospheric light and the transmission map based on reasonable assumptions [10][11][12][13]. He et al. proposed a dark channel prior (DCP)-based haze removal method [14]. They assumed that pixels in the local patch of a clear image have at least one dark pixel. The DCP method works well in most regions that satisfy the DCP assumption, but fails in a white object region. Berman et al. estimated the transmission map using haze-line prior assumption that the pixel coordinates in the color space tend to become closer to the atmospheric light in a hazy image [15]. To find the lower bound of a haze-line, they used the 500 representative colors. While the Berman's approach enhances color contrast, it is impossible to find representative colors in a severely degraded image by haze or fog. Shin et al. optimized the transmission estimation process using both radiance and reflectance components [16].
Recently, convolutional neural networks (CNN) are being applied not only to image classification, but also to variety of low-level image processing applications [17][18][19][20]. The CNN-based dehazing methods were also proposed in the literature to overcome the limitation of the transmission map estimation using a single image. Cai et al. estimated the transmission to restore a haze image using a DehazeNet [21]. Cai's method falls in the end-to-end supervised learning approach using synthetic haze and clean patches. To overcome the limitation of haze feature estimation, Ren et al. presented a multi-scale CNN [22]. They also proposed a learning method using the pairs of the simulated haze image and true transmission [23].
To increase the training accuracy, Li et al. combined two CNN modules of the transmission and atmospheric light estimation via all-in-one dehazing network (AODNet) [24]. Zhang et al. proposed a densely connected pyramid dehazing network (DCPDN) optimized by a conditional adversarial learning method [25,26]. The depth information can be incorporated into the transmission estimation process using a supervised learning method. However, it is hard to reflect other quantities such as attenuation, atmospheric light, and illuminance at once because it is difficult to collect the data including the depth, attenuation, airlight, and ideal illuminance maps.
For example, Figure 1a shows a real-haze image provided by [27]. This type of haze in Figure 1a is different from what we have simulated, and degraded by multiple factors including the color attenuation, unbalanced light source and scattered light. Therefore, CNN-based estimation can not adaptively remove this real-haze as shown Figure 1b Note that the proposed method can restore the most naturally looking image by removing real-haze based on the direct estimation of the radiance map.
To overcome the dependency, a radiance estimation method can be applied to the dehazing process. Ren et al. estimated the haze-free radiance by using a mult-scale convolutional neural network and simulated haze dataset [22]. The mult-scale convolutional neural network can stably remove the simulated haze. Chen et al. estimated a physical haze modelbased radiance image using a dilated convolution [18] and adaptive normalization [28]. It can approximate the DCP or non-local dehazing operators using low computational complexity. This radiance estimation method can accurately estimate the dehazed result without additional estimation steps, but it may generate the amplified noise and dehazing artifacts. To approach fusion method, Ren et al. removed the haze using derived inputs and gated fusion network [29], Shin et al. proposed the triple convolutional networks including dehazing, enhancement, and concatenating subnetworks to enhance the contrast without dehazing artifacts [30]. However, the separated subnetworks result in increasing computational complexity. To solve this problem, this paper presents a new dehazing and verifying network (DVNet). The proposed DVNet does not need the subnetworks in the prediction procedure. Instead, only correction subnetwork is used for the training process, and evaluates the dehazing error in the output using a complementary adversarial learning. Different from the transmission estimation-based method, the proposed DVNet successfully removed the real-haze without the noise, halo, or other undesired artifacts with low computational complexity. Since the proposed method can use more enhanced ground truth images, our DVNet can be effectively learned by using absolute-mean error and perceptual loss functions. Furthermore, our verifying network simultaneously estimates and reduces the error of the resulting images via self supervised learning and least square adversarial network. Therefore, experimental results show that the proposed DVNet outperforms existing state-of-the-art approaches in the sense of both robustness to various haze environment and computational efficiency. This paper is organized as follows: Section 2 summarized related works, and Section 3, respectively, describes the proposed DVNet and the corresponding training method. After summarizing experimental results in Section 4, we conclude the paper with some discussions in Section 5.
Related Works
A clear image is degraded by the physical haze model as [5] x where J represents a haze-free, clean image, x the hazy, degraded version, p the twodimensional pixel coordinate, t the light transmission map, and A the spatially-invariant atmospheric light. Superscripts in x, J, and A represent a color channel, and the transmission t(p) is independent of the color channel. To solve this equation, physical haze model-based methods estimate the major components such as t and A based on a proper assumptions. Recently, several deep learning techniques can make this formula solvable without estimating t or A estimations. Therefore, this section introduces various deep learning-based dehazing approaches.
Physical Haze Model-Based Dehazing
He et al. applied the dark channel prior (DCP) to estimate the transmission as [14] t DCP (p) = 1 − min where q is the 2D pixel coordinate in a local patch region around p, denoted as N (p), in which the transmission is assumed to be constant. Berman et al. estimated the non-local (NL) transmission map using the geometric haze feature as [15] To solve for the feature in (3), Berman et al. used 500 representative colors and approximated the denominator using the k-nearest neighbor (k-NN) algorithm [31]. To minimize the dehazing artifacts such as noise and halo in the estimated transmission, either soft matting or weighted least squares [32,33] algorithm can be used as a regularization function. Shin et al. estimate the transmission by minimizing the radiance-reflectance combined cost as [16] arg min
Radiance-Based Dehazing
Given N pairs of haze-free and its hazy version patches, CNN-based dehazing methods commonly train the network by minimizing the loss function as where J P i and x P i represent the i-th training patches of the haze-free and hazy images, respectively. Θ is a set of network parameters including weights and biases, and F (·) is the output of the network given an input hazy image patch and the set of parameters [28,34].
Adversarial Learning
To reduce the divergence between the generated and real images, the adversarial loss can be defined as [26,[35][36][37] arg min where G J is the haze-free generator, D is a discriminator to discriminate a real or fake class, and L{·} denotes a sigmoid cross entropy operator. This adversarial learning can generate a haze-free image that is closer to the clean image.
Proposed Method
To remove haze, we present a new dehazing and verifying networks using dilated convolution layers and generative adversarial network. Deep learning-based dehazing methods require a serious of procedures including: Generation of dataset, configuration of a deep learning model, and training the model. In this section, we describe the data generation method in Section 3.1, the network architecture and learning functions of both correction and haze nets are given in Sections 3.2 and 3.3. Section 3.4 presents the proposed training approaches including the verifying network and complementary adversarial learning.
Data Generation
To generate the pairs of the haze and clean images, we first generate the initial dehazed image from the input hazy image using a physical haze model given in (1). Let I(p) be the input hazy image, andt(p) the estimated transmission using either (3) or (4), the initial clean image is computed as Since (7) gives an one-step, closed-form estimation, the training pairs of the hazy and haze-free images can be easily created. In this paper, we used the result of the non-local dehazing (NL) and radiance-reflectance optimization method (RRO) given in (3) and (4) to generate the initial dehazed images. In addition, haze simulated images such as NYUdepth data [23] can also be used to generate I D and I in pair based on physical haze-model. Overall, the generated data I D is used to input data of the correction network as shown in Figure 2. In the dehazing procedure, the input haze images are resoted by the haze network, which is learned by the corrected images. The verifying network imitates the natural images using self supervised learning, and the discriminator classifies the real or fake class between the natural image and generated images to reduce the statistical divergence.
Correction-Network (CNet)
We propose a correction network (CNet) to enhance the initial dehazed images by correcting both color and intensity values. To restore the missing information, we concatenate features of the haze network (HNet) using the dilated convolution and adaptive normalization [18,28] wheref k i and b i k , respectively, represent the i-th feature map and bias in the k-th layer, and is the kernels to obtain the i-th feature map using the feature maps extracted in the The operator " * r k " represents the dilated convolution using the rate of the k-th layer, r k . The dilated convolution can quickly perform filtering in a wide receptive field without changing the scale. g is a leaky rectified linear unit (LReLU) [38] function defined as A k (·) represents the adaptive normalization (AN) function in the k-th layer as where BN(·) denotes the batch normalization function [39], α k and β k are the trainable parameters to control the relative portion of the batch normalization function. The adaptive normalization approach given in (10) can provide an enhanced restoration results [28].
is concatenated as where concat is a feature concatenation operator [40], f is the feature map in a HNet that will be described in Section 3.3. This connection plays an important role in coordinating the learning direction. For example, if the CNet is incorrectly learned without the upward connections, the HNet is also learned with different images and such erroneous cycles are repeated. To correctly propagate the learning direction, we concatenate the feature maps of the HNet to the upward feature maps of the CNet. Top of Figure 2 shows the CNet and the proposed upward connection scheme. In addition, the parameters of CNet can be optimized by self-supervised learning using the perceptual loss [41], and it can be defined by VGG16 network [42] which is pretrained using ImageNet data [43]. The perceptual loss in the CNet is referred to as correction loss, which is defined as where N represents the batch size, I C the output of the C-Net, and F returns the feature maps of the VGG16 network model. We used relu1-2, relu2-2, relu3-3 and relu4-3 features in the VGG16. λ is a parameter to regularize 1 -norm of the gradient. This self-supervised CNet can correct color, intensity, and saturation in real-hazy dataset [27] as shown in Figure 3.
Haze Images
Initial Dehazed Images Corrected Images (CNet)
Haze-Network (HNet)
The HNet plays an important role in enhancing the degraded images. In addition, an efficient design of the H-Net can significantly reduce the processing time. For that reason, the HNet uses the dilated convolution and adaptive normalization [18,28] as, where f i k is a feature map of the H-Net in the k-th layer. b, h, and A k (·), respectively, represent the bias, kernel and adaptive normalization operator. Since the HNet is learned using the results of the CNet, its result can also be corrected in an adaptive manner. The HNet can be optimized by minimizing the haze loss as: where I H i is the output of the HNet.
Verifying Network
To make the outputs of the dehazing network (HNet, CNet) look more natural, we verify the errors, such as noise and halo artifact, using self-supervised learning with clean data [44]. The verifying loss of the self-supervised learning is defined as where I N i , IV i , and I V i , respectively, represent the clean image, results of the CNet, and HNet. Note that the self-supervised terms are designed by considering the errors, which means that the pixels and features in output images of both CNet and HNet are closed to the real natural images when the input images are ideally clean [30]. If input images are the clean images, the ideal haze model should generate the same natural images as in the left-bottom of Figure 2. Therefore this self supervised loss should be separately applied to optimize the networks as Algorithms 1 and 2. In this context, the self-supervised learning based on the loss in (15) using a clean image can minimize the dehazing artifacts as shown in Figure 4d. Futhermore, to reduce the statistical divergence between the generated and real images, the proposed DVNet can be optimized based on the least square adversarial cost [36] and min where D is a convolutional neural net based dicriminator as shown in right-bottom of Figure 2, which returns a probablity value of the input image I * using a binary softmax algorithm. G is the generative networks including HNet and CNet. The input data of the discriminator is the ideally natural data I N , and the random noise is replaced to real-haze image I in , the initial dehazed image I D , and natural image I N to engage our HNet and CNet.
In this adversarial learning method, the proposed network can be learned to reduce the probability divergence between the clean image I N and the result of the proposed network (I H , I C , I V ) using unfair images. To implement the adversarial cost, we will describe about the optimal parameters in Appendix A.
Therefore, the resulting images (I H , I C , I V ) can be improved as the visibility is similar to the clean images (I N ). Figure 4e shows the performance of the proposed DVNet. More specifically, the resulting images in Figure 4 show that our DVNet can better enhance the hazy images [45] in the sense of both details and contrast without the undesired dehazing artifacts.
Implementation
For the implementation, we split our method into the training and testing procedures. The training procedure consists of eight steps: (i) Feature extraction using HNet, (ii) feature concatenation using the CNet and generation of the corrected clean image, (iii) error verification using the same network architecture and natural image [44], (iv) differentiation of the real and fake images using discriminator, (v) minimizing (14) + (12), (vi) minimizing (15), (vii) maximizing and minimizing adversarial costs V(D) and V(G), (viii) repeat the above seven steps until the optimal CNN weights are obtained. The test procedure is simpler than the training procedure, and applies the optimal HNet to remove haze. Table 1 shows the pseudo-code of training and testing procedures of the proposed method. In Tables 2 and 3, the parameters of the proposed DVNet and discriminator are given for the implementation. To optimize the cost functions, we used an adaptive moment estimation (ADAM) optimization algorithm proposed by [46]. Learning rate values of the DNet and VNet were, respectively, set to 1 × 10 4 and 4 × 10 4 . We used 500 real-haze images from the dataset provided by [27], which are engaged to the DVNet with high quality images from NITRE 2017 dataset [44]. Initial clean images were created using the NL, RRO, and NYU-depth data [15,16,23] using five hundred training images. We trained the proposed DVNet 10,000 times. Table 1 shows conventions for the important variables and parameters for the implementation.
Experimental Results
For the experiment, we selected three benchmark datasets of size 512 × 512 including I-Haze, O-Haze, and 100 real hazy images [27,[47][48][49]. Especially for the comparative experiment, we tested existing dehazing methods including: Haze-line prior-based nonlocal dehazing method (NL), densely connected pyramid dehazing net (DCPDN), radiancereflectance optimization based dehazing (RRO), the region-based haze image enhancement method by using triple convolution network(TCN) [15,16,25,30]. Both NL and RRO were implemented in Matlab 2016b and tested on i7 CPU equipped with 64 GB of RAM. On the other hand, DCPDN, TCN and the proposed method were tested using NVIDIA RTX 2080ti graphics processing unit (GPU) and implemented in Python version 3.6 and Tensorflow. This section includes similarity evaluation in Section 4.1, visual quality evaluation in Section 4.2, and ablation study in Section 4.3.
For the quantitative evaluation, we measured the peak signal to noise ratio (PSNR), structural similarity index measure (SSIM), and CIE color difference formula 2000 (CIED) [50,51] as shown in Figures 5 and 6 and Table 4, where the best and second best scores are, respectively, shown with blue and cyan colored text. The proposed DVNet is trained by non-local dehazing or radiacne-reflectance optimization-based restoration results or NYU-depth dataset based haze-clean pair. Both DVNet-RRO and DVNet-NL outperform than state-of-the-art approaches in term of both SSIM, and CIED in I-Haze dataset, which has the ideal illumination because each image was acquired in the indoor environment. However, the performance of DVNet-NYU was slightly lower than TCN-RRO in term of PSNR and SSIM because simulated dataset can not reflect various environments such as airlight and illuminance. It means that the DVNet-NYU can generate intensity saturation as shown in Figure 5h.
Since adaptive normalization used in the TCN and our DVNet stretches the intensity, both DVNet and TCN can change the background color. Therefore, the PSNR of the DVNet-RRO is similar to that of TCN. Note that the DVNet does not only remove the haze but also change the illumination. So the resulting image has a different illuminance from the ground truth image. For that reason, the DVNets and TCN produce a lower similarity in the O-Haze dataset than the NL and RRO approaches.
However the DVNet-RRO performs better than other CNN-based methods such as DCPDN and TCN in term of SSIM.
Visual Qaulity Assessment
To verify the performance of the DVNets in the real haze conditions, we used 100-FADE test sets provided by [27]. For the objective evaluation, we select no-reference measures including: Contrast to noise ratio (CNR), natural image quality evaluation (NIQE), entropy to evaluate amount of information in a single image such as intensity distribution, and intensity saturation [27,52,53]. A high-quality image has high CNR and entropy values, whereas it should have a low NIQE and saturation values for stable enhancement. The average scores of the proposed DVNet-NL are higher than those of stateof-the-art approaches in terms of the CNR and saturation as shown in Table 5. The ranking of the DVNet-NYU was the best score in terms of CNR, entropy, and NIQE. However, due to highly saturated pixels, the color of resultant image of DVNet-NYU can be distorted as shown in Figure 7h. Note that the DVNet-NL has high score in terms of the NIQE with a very small difference from the first NL. The DVNet-RRO also has a similar score in term of NIQE compared with RRO. However, the saturation score of the DVNets are lower than NL and RRO because our DVNets verifies the errors of the NL, RRO, and NYU-depth dataset. In summary, the proposed DVNet can successfully remove various types of haze in various environment [27] as shown in Figures 7 and 8.
Additional Study
To demonstrate the effect the proposed contributions, we conducted the additional studies using the I-Haze and O-Haze datasets. We also used version of the DVNet-NL for the ablation study. In Table 6, HNet and CNet represent the baseline of the proposed dehazing network, DVNet the optimized version of the proposed method with the natural image and self-supervised learning, GAN the optimized version of the proposed method using the proposed adversarial learning method. Note that the combined HNet and CNet model without VNet returns only similar images to those of physical model-based dehazing method, which also imitates the error such as noise and saturation. Our DNet (HNet + CNet) can reduce the intensity distortion caused by initial dehazed image I D . The SSIM values the DVNet increased at the cost of a slight PSNR reduction. This means that our verifying process can prevent the noise and halo at the cost of slightly reduced dehazing performance. However, since the proposed adversarial network complements the dehazing performance, the PSNR values outperform the vanilla DVNet. In addition, Table 7 shows the processing time of the proposed DVNet with various image sizes. In evaluation procedure, the proposed DVNets only use a single network(HNet). Therefore, the DVNets can more reduce the computational time over 5-10 times than the TCN and DCPDN, which have several subnetworks.
Conclusions
To estimate a high-quality, clean radiance image without the dehazing artifacts, we proposed a novel dehazing network followed by a verifying network, which generates the radiance images to verify the dehazing errors. To estimate an ideally clean image pair, we concatenate feature maps using adaptive normalization and upward connections from the HNet to the CNet. In addition, an unpaired natural image and the discriminator can help minimizing the noise and dehazing artifacts without the performance degradation. The DVNet can be adaptively remove the haze without addtional estimation processes. Therefore, the proposed approach can efficiently remove various types of haze with low conputational complexity. More specifically, three experiments were conducted to verify the performance of the DVNet and the effect of the individual contributions. As a result, the DVNet can provide high-quality dehazing results under various types of haze environments. However, the DVNet may depend on the based training data. In the future work, we plan to combine the DVNet with the data augmentation method, and expand it to video dehazing.
Conflicts of Interest:
The authors declare no conflict of interest.
Appendix A. Optimal Parameters
In proposed method, the least square adversarial cost functions are defined as [36] max where D returns the probability values via the discriminator using the soft-max algorithm, and G represents the proposed generator model including H and CNet. To find the optimal point of the discriminator, V(D) in (A1) can be expressed as The optimal point of the discriminator D * can be obtained when its partial derivative with respect to D is equal to zero, such as Therefore, the optimal point D * can be defined as which can be simplified by defining the real and fake distributions, respectively, denoted as P 1 = P D and P 2 = 3P G , D * (x) = bP 1 + aP 2 P 1 + P 2 , (A7) (A2) is expressed as and V(G) = x ((b−c)(p 1 (x)+p 2 (x))−(b−a)p 2 (x)) 2 p 1 +p 2 dx.
If we set conditions as: b − c = 1, b − a = 4 3 , and P 1 ≈ 1 3 P 2 , then V(G) will converge. Therefore, (A10) is re-written as V(G) = x 4 3 P 2 (x) − (P 1 (x) + P 2 (x)) 2 P 1 + P 2 dx, (A12) where χ 2 P represents Pearson-χ 2 divergence [36]. It means that when the above conditions are satisfied, χ 2 divergence can minimize the distance between P 1 + P 2 and 4 3 P 2 . So, above equation can be expressed as If all conditions are satisfied, then P D = P G . Therefore, the optimal parameters can be defined as a = 4/3, b = 0, and c = 1. However, since the maximum value of D is equal to 1, the proposed parameters are applied as | 6,039.6 | 2021-09-01T00:00:00.000 | [
"Computer Science"
] |
Study of Intra-Chamber Processes in Solid Rocket Motors by Fiber Optic Sensors
In this study, an experimental study of the burning rate of solid fuel in a model solid propellant rocket motor (SRM) E-5-0 was conducted using a non-invasive control method with fiber-optic sensors (FOSs). Three sensors based on the Mach–Zehnder interferometer (MZI), fixed on the SRM E-5-0, recorded the vibration signal during the entire cycle of solid fuel burning. The results showed that, when using MZI sensors, the non-invasive control of solid fuel burnout is made possible both by recording the time of arrival of the combustion front to the sensor and by analyzing the peaks on the spectrogram of the recorded FOS signal. The main mode of acoustic vibrations of the chamber of the model SRM is longitudinal, and it changes with time, depending on the chamber length. Longitudinal modes of the combustion chamber were detected by MZI only after the combustion front passed its fixing point, and the microphone was unable to register them at all. The results showed that the combustion rate was practically constant after the first second, which was confirmed by the graph of the pressure versus time at the nozzle exit.
Introduction
In modern engines and, particularly, in a solid propellant rocket motor (SRM), energy conversion processes are characterized by extreme temperatures and released power. For example, the thermodynamic temperature is around 3600 K in the combustion chamber of the Ariane-V launch vehicle's accelerator, EAP P241, which produces a thrust of 7.08 MN [1]. Under such conditions, the SRM operation parameters change frequently due to rates of the main intra-chamber processes-for example, fuel combustion. Therefore, we can describe the working process parameters in SRM as fast-flowing. They include vibrations, pressure in the combustion chamber, and acousto-optical and electrophysical characteristics [2,3].
There are various ways to monitor these parameters. Some methods include rather complex installations using, for example, X-ray analysis [4][5][6] for intra-chamber process control. This method allows for the observation of the fuel mass combustion patterns through the walls of the housing [7]. However, the sampling frequency is low, and the complexity of the setup makes this method inaccessible for common usage. There are also methods for the invasive monitoring of such processes, for example, by inserting thermocouples into test holes. This is simpler, but violates the integrity of the engine and probably changes the operation's parameters. Additionally, the connection wires require a lot of space, add mass to the setup, and can produce sparks. If we place thermocouples on the motor casing in a non-invasive way, the measurement obtained will be incorrect
Theory
Pressure, temperature, deformation, and vibration constitute the most important parameters of the working process in a model SRM chamber. Currently, only a limited number of FOSs are able detect these values [28,40,41]. The available data suggests the following requirement: an expected total measurement time of less than 10 seconds; typical oscillation frequencies of intra-chamber processes starting from tens of Hz, increasing to hundreds of kHz. In this case, optical time-domain reflectometry (OTDR) systems and devices based on FBG are insufficient due to the low sampling frequency of 30 kHz for phi-OTDR [42][43][44] and 10 kHz for FBG in configurations where there are a few sensors in one line [45][46][47]. In turn, the small SRM length and the requirement of less sensing points allows for the application of sensors based on the MZI. A phase-unwrapping technique was used for dynamic range improvement, based on the 3 × 3 output splitter providing phase-shifted signals.
A fiber MZI registers a phase difference between two arms, one of which is a reference and the other a sensing arm, as shown in Figure 1. No load is applied to the reference arm during the experiment. The sensing arm is fixed on the controlled object, and its length changes depending on the object's deformation. This leads to a signal intensity fluctuation, which is proportional to the cosine function of deformation. A 3 × 3 output splitter produces a 2π/3 phase shift between neighboring fibers. Thus, the recorded intensity on each photoreceiver can be determined by the following expression [48]: where I 1 , I 2 are the intensities from the reference and sensing arms, respectively; λ is the wavelength of laser radiation, m; ∆(t) is the optical path difference between the MZI arms, m and ϕ 0 is the initial phase difference, rad. A fiber MZI registers a phase difference between two arms, one of which is a reference and the other a sensing arm, as shown in Figure 1. No load is applied to the reference arm during the experiment. The sensing arm is fixed on the controlled object, and its length changes depending on the object's deformation. This leads to a signal intensity fluctuation, which is proportional to the cosine function of deformation. A 3 × 3 output splitter produces a 2π/3 phase shift between neighboring fibers. Thus, the recorded intensity on each photoreceiver can be determined by the following expression [48]: where I1, I2 are the intensities from the reference and sensing arms, respectively; λ is the wavelength of laser radiation, m; Δ(t) is the optical path difference between the MZI arms, m and φ0 is the initial phase difference, rad. The presence of two or more signals with a 2π/3 phase shift allows for the restoration of the phase φ of the deformation without uncertainty caused by a cosine function. We implemented a scheme with two photoreceivers for each MZI, reduced the number of photoreceivers, and simplified the measurement scheme. The deformation phase change Δϕ was obtained by the algorithm described in [49]. In this case, at each time t, it was calculated by the formula: where ( ) = ( ) − ( ), ( ) = ( ) + ( ).
The phase change depends on the deformation of the fiber, which was influenced by thermal, mechanical, and acoustic effects. In the experiment, the fiber was coiled in certain places on the outer surface of a model SRM, as shown in Figures 1 and 3. The total sensitive fiber length was Lsens = Nπd, where d = 19 mm is the initial outer diameter of the SRM housing and N = 10 is the number of turns. The SRM diameter increase, and phase change are linked through the fluctuations of the sensing arm length ΔLsens by the formula: where n is the effective refractive index of the fiber core, and consequently, The presence of two or more signals with a 2π/3 phase shift allows for the restoration of the phase ϕ of the deformation without uncertainty caused by a cosine function. We implemented a scheme with two photoreceivers for each MZI, reduced the number of photoreceivers, and simplified the measurement scheme. The deformation phase change ∆φ was obtained by the algorithm described in [49]. In this case, at each time t, it was calculated by the formula: where S 1 (t) = I PD1 (t) − I PD2 (t), S 2 (t) = I PD1 (t) + I PD2 (t).
The phase change depends on the deformation of the fiber, which was influenced by thermal, mechanical, and acoustic effects. In the experiment, the fiber was coiled in certain places on the outer surface of a model SRM, as shown in Figures 1 and 3. The total sensitive fiber length was L sens = Nπd, where d = 19 mm is the initial outer diameter of the SRM housing and N = 10 is the number of turns. The SRM diameter increase, and phase change are linked through the fluctuations of the sensing arm length ∆L sens by the formula: where n is the effective refractive index of the fiber core, and consequently, A frequency analysis of the phase change provides additional information about the combustion process. The SRM chamber's acoustic vibration modes depend on its size. The main types are the first longitudinal f lon , tangential f tan , and radial f rad , determined by the formula [50][51][52]: where a is the speed of sound, m/s, and l(t) is the combustion chamber length, m, at time t, s, as shown in Figure 2. Length l(t) varied from 8 to 109 mm during the experiments; d i = 15 mm is the chamber internal diameter. A frequency analysis of the phase change provides additional information about the combustion process. The SRM chamber's acoustic vibration modes depend on its size.
The main types are the first longitudinal flon, tangential ftan, and radial frad, determined by the formula [50][51][52]: where a is the speed of sound, m/s, and l(t) is the combustion chamber length, m, at time t, s, as shown in Figure 2. Length l(t) varied from 8 to 109 mm during the experiments; di = 15 mm is the chamber internal diameter. These frequencies contribute to phase change fluctuations and can be observed in the spectra. This observation method can precisely determine the chamber length.
Description of the Experimental Setup and Methods for Recording the Characteristics of Intra-Chamber Processes
In this study, a model SRM E-5-0 is the research object. It operates via a black powder [53], that was pressed into a cylindrical body made of cardboard. A graphite nozzle block was installed on the bottom with a critical section diameter of 3.4 mm. Fuel ignition was performed using a 0.5 g black powder sample via a combustible wire. The model SRM characteristics are shown in Table 1. Figure 3 shows the SRM photograph (a), a diagram with dimensions between the main components and the MZIs (b), and a section of the SRM after the study was conducted (c). The experimental setup included three MZIs; its scheme is shown in Figure 4. The MZI sensing arms were fixed equidistantly along the entire fuel length. These frequencies contribute to phase change fluctuations and can be observed in the spectra. This observation method can precisely determine the chamber length.
Description of the Experimental Setup and Methods for Recording the Characteristics of Intra-Chamber Processes
In this study, a model SRM E-5-0 is the research object. It operates via a black powder [53], that was pressed into a cylindrical body made of cardboard. A graphite nozzle block was installed on the bottom with a critical section diameter of 3.4 mm. Fuel ignition was performed using a 0.5 g black powder sample via a combustible wire. The model SRM characteristics are shown in Table 1. Figure 3 shows the SRM photograph (a), a diagram with dimensions between the main components and the MZIs (b), and a section of the SRM after the study was conducted (c). The experimental setup included three MZIs; its scheme is shown in Figure 4. The MZI sensing arms were fixed equidistantly along the entire fuel length. A narrow-band NKT BASIK MIKRO fiber laser with a central wavelength of 1550 nm and a bandwidth of less than 0.1 kHz was used. Its radiation emitted through a 3 × 3 splitter to three independent equal-arm MZIs. The supporting arms were at rest, and the sensing arms were coiled on the SRM housing (see Figure 1). The measuring arm of each A narrow-band NKT BASIK MIKRO fiber laser with a central wavelength of 1550 nm and a bandwidth of less than 0.1 kHz was used. Its radiation emitted through a 3 × 3 splitter to three independent equal-arm MZIs. The supporting arms were at rest, and the sensing arms were coiled on the SRM housing (see Figure 1). The measuring arm of each A narrow-band NKT BASIK MIKRO fiber laser with a central wavelength of 1550 nm and a bandwidth of less than 0.1 kHz was used. Its radiation emitted through a 3 × 3 splitter to three independent equal-arm MZIs. The supporting arms were at rest, and the sensing arms were coiled on the SRM housing (see Figure 1). The measuring arm of each MZI consisted of L sens = Nπd = 10·π·19 mm ≈ 0.6 m of SMF-28. We glued this fiber loop-to-loop using one layer of double-sided tape. This method of construction increased the MZI sensitivity to fluctuations of housing diameter [54]. The SRM was fixed on the metal table by clamps. Such mounting proved adequate to complete the measurements; the motor shifted slightly at the start as a result of the highest pressure, and the sensing fiber remained connected to the housing at all times. The process was recorded, and is provided in the attached Video S1. Two fibers of each MZI 3 × 3 splitter outputs were transferred to photodiodes (PD). The signals were digitized on an ADC with a sampling rate of 2.5 MHz. This value determined the maximum detectable vibration frequency (1.25 MHz), according to the Nyquist theorem. An image of a laboratory setup with the measurement and registration systems is shown in Figure 5. Before the experiment, we checked the setup's integrity and its ability of deformation registration.
Sensors 2021, 21, x FOR PEER REVIEW MZI consisted of Lsens = Nπd = 10·π·19 mm ≈ 0.6 m of SMF-28. We glued thi loop-to-loop using one layer of double-sided tape. This method of construction inc the MZI sensitivity to fluctuations of housing diameter [54]. The SRM was fixed metal table by clamps. Such mounting proved adequate to complete the measure the motor shifted slightly at the start as a result of the highest pressure, and the s fiber remained connected to the housing at all times. The process was recorded, provided in the attached Video S1. Two fibers of each MZI 3 × 3 splitter output transferred to photodiodes (PD). The signals were digitized on an ADC with a sam rate of 2.5 MHz. This value determined the maximum detectable vibration freq (1.25 MHz), according to the Nyquist theorem. An image of a laboratory setup w measurement and registration systems is shown in Figure 5. Before the experime checked the setup's integrity and its ability of deformation registration.
Analysis of the Investigation Results
Images of the SRM stages are shown in Figure 6, including start-up (a,b), ope in nominal mode (c,d), and shutdown (e,f). It is worth noting that the tracks of th densed phase particles flowed out of the engine nozzle, which is common for the bustion products of powder and metal-containing fuels. The total operating tim about 5.5 s, during which the optical fiber did not undergo any damage or changes the effect of high-temperature combustion products.
Analysis of the Investigation Results
Images of the SRM stages are shown in Figure 6, including start-up (a,b), operation in nominal mode (c,d), and shutdown (e,f). It is worth noting that the tracks of the condensed phase particles flowed out of the engine nozzle, which is common for the combustion products of powder and metal-containing fuels. The total operating time was about 5.5 s, during which the optical fiber did not undergo any damage or changes due to the effect of high-temperature combustion products. MZI consisted of Lsens = Nπd = 10·π·19 mm ≈ 0.6 m of SMF-28. We glued this fiber loop-to-loop using one layer of double-sided tape. This method of construction increased the MZI sensitivity to fluctuations of housing diameter [54]. The SRM was fixed on the metal table by clamps. Such mounting proved adequate to complete the measurements; the motor shifted slightly at the start as a result of the highest pressure, and the sensing fiber remained connected to the housing at all times. The process was recorded, and is provided in the attached Video S1. Two fibers of each MZI 3 × 3 splitter outputs were transferred to photodiodes (PD). The signals were digitized on an ADC with a sampling rate of 2.5 MHz. This value determined the maximum detectable vibration frequency (1.25 MHz), according to the Nyquist theorem. An image of a laboratory setup with the measurement and registration systems is shown in Figure 5. Before the experiment, we checked the setup's integrity and its ability of deformation registration.
Analysis of the Investigation Results
Images of the SRM stages are shown in Figure 6, including start-up (a,b), operation in nominal mode (c,d), and shutdown (e,f). It is worth noting that the tracks of the condensed phase particles flowed out of the engine nozzle, which is common for the combustion products of powder and metal-containing fuels. The total operating time was about 5.5 s, during which the optical fiber did not undergo any damage or changes due to the effect of high-temperature combustion products. The recorded data from each MZI were processed in the time and frequency domains. The data ranges from 2 seconds before fuel ignition through the combustion process to around 2 seconds after its completion. Until the engine was turned on, the signal The recorded data from each MZI were processed in the time and frequency domains. The data ranges from 2 seconds before fuel ignition through the combustion process to around 2 seconds after its completion. Until the engine was turned on, the signal at each PD changed with a small amplitude. The high-frequency component occurred due to the PD and the laser phase noise, and the low-frequency fluctuation was a result of the installation temperature drift and the laser wavelength drift. At launch, the amplitude increased in signal oscillations on all PDs. An example of the initial data from one channel of each MZI is presented in Figure 7-the oscillation amplitude increased on all interferometers from the moment the engine was launched, but it only reached the maximum contrast when the combustion surface of the solid fuel reached the MZI sensing arm on the SRM housing.
The recorded data from each MZI were processed in the time and freque mains. The data ranges from 2 seconds before fuel ignition through the combusti cess to around 2 seconds after its completion. Until the engine was turned on, th at each PD changed with a small amplitude. The high-frequency component o due to the PD and the laser phase noise, and the low-frequency fluctuation was a r the installation temperature drift and the laser wavelength drift. At launch, the tude increased in signal oscillations on all PDs. An example of the initial data fr channel of each MZI is presented in Figure 7-the oscillation amplitude increase interferometers from the moment the engine was launched, but it only reac maximum contrast when the combustion surface of the solid fuel reached the M ing arm on the SRM housing. For each sensor, the phase-unwrapping procedure was carried out accor Formula (2). An absolute value of the optical signal phase change, from the init (before the engine launch), was obtained and was found to be proportional to the of the fiber length on the model SRM according to Equation (3). The results highli the closer the MZI to the nozzle, the more changes it experienced. The plots housing diameter increase are presented in Figure 8. For each sensor, the phase-unwrapping procedure was carried out according to Formula (2). An absolute value of the optical signal phase change, from the initial state (before the engine launch), was obtained and was found to be proportional to the change of the fiber length on the model SRM according to Equation (3). The results highlight that the closer the MZI to the nozzle, the more changes it experienced. The plots for the housing diameter increase are presented in Figure 8. The derivatives of the housing diameter expansion graphs were calculated with a 60 ms window, allowing for the exclusion of high-frequency oscillations and their influence on the derivative stability. All of the sensors had a moment of initial expansion at the engine start, after which the diameter value became relatively stable, without a noticeable trend of expansion. Graphs illustrating the derivatives from each sensor are shown in Figure 9. A sharp increase in the derivative was observed when the combustion The derivatives of the housing diameter expansion graphs were calculated with a 60 ms window, allowing for the exclusion of high-frequency oscillations and their influence on the derivative stability. All of the sensors had a moment of initial expansion at the engine start, after which the diameter value became relatively stable, without a noticeable trend of expansion. Graphs illustrating the derivatives from each sensor are shown in Figure 9. A sharp increase in the derivative was observed when the combustion surface coordinate reached the sensor fixing point. These points are marked with circles in Figure 9. The derivatives of the housing diameter expansion graphs were calculated 60 ms window, allowing for the exclusion of high-frequency oscillations and thei ence on the derivative stability. All of the sensors had a moment of initial expan the engine start, after which the diameter value became relatively stable, withou ticeable trend of expansion. Graphs illustrating the derivatives from each sen shown in Figure 9. A sharp increase in the derivative was observed when the com surface coordinate reached the sensor fixing point. These points are marked with in Figure 9. The coordinates of the combustion-front propagation were determined by t of derivative sharp growth for the sensors, and are shown in Table 2. Based o values, we graphed the combustion surface movement, as presented in Figure 10. The coordinates of the combustion-front propagation were determined by the time of derivative sharp growth for the sensors, and are shown in Table 2. Based on these values, we graphed the combustion surface movement, as presented in Figure 10. The burning rate is non-linear in the first and smaller section of the graph due to the uneven combustion front, caused by the presence of a groove at the end of the solid fuel, as well as the combustion of the igniter sample, also made of black powder. After the end of the ignition period, the time dependence of the combustion surface movement was found to be close to linear with an average linear displacement velocity of 0.0193 m/s. This dependence (nonlinear during the ~1 second and then linear) is consistent with the results of a similar SRM test, showing that the pressure in the combustion chamber after around 1 second, following the engine start, became almost constant, as shown in Figure 11. The pressure in the combustion chamber during the experiment was measured using a special setup. The SRM was installed in a stainless steel external chamber with a pressure sensor. This setup was the only method by which to fix the pressure sensor to The burning rate is non-linear in the first and smaller section of the graph due to the uneven combustion front, caused by the presence of a groove at the end of the solid fuel, as well as the combustion of the igniter sample, also made of black powder. After the end of the ignition period, the time dependence of the combustion surface movement was found to be close to linear with an average linear displacement velocity of 0.0193 m/s. This dependence (nonlinear during the~1 second and then linear) is consistent with the results of a similar SRM test, showing that the pressure in the combustion chamber after around 1 second, following the engine start, became almost constant, as shown in Figure 11. The pressure in the combustion chamber during the experiment was measured using a special setup. The SRM was installed in a stainless steel external chamber with a pressure sensor. This setup was the only method by which to fix the pressure sensor to SRM. The described modification slightly increased the combustion time to 6.5 s. However, in general, the pressure change during the investigation remained unchanged for all SRMs of such a model. The fuel and housing construction provide a constant combustion surface area for when the fuel burns, therefore, a constant pressure in the chamber after~1 second after start becomes apparent even in the presence of deviations in the initial temperature, solid fuel composition, critical section diameter, etc.
The burning rate is non-linear in the first and smaller section of the graph due to the uneven combustion front, caused by the presence of a groove at the end of the solid fuel, as well as the combustion of the igniter sample, also made of black powder. After the end of the ignition period, the time dependence of the combustion surface movement was found to be close to linear with an average linear displacement velocity of 0.0193 m/s. This dependence (nonlinear during the ~1 second and then linear) is consistent with the results of a similar SRM test, showing that the pressure in the combustion chamber after around 1 second, following the engine start, became almost constant, as shown in Figure 11. The pressure in the combustion chamber during the experiment was measured using a special setup. The SRM was installed in a stainless steel external chamber with a pressure sensor. This setup was the only method by which to fix the pressure sensor to SRM. The described modification slightly increased the combustion time to 6.5 s. However, in general, the pressure change during the investigation remained unchanged for all SRMs of such a model. The fuel and housing construction provide a constant combustion surface area for when the fuel burns, therefore, a constant pressure in the chamber after ~1 second after start becomes apparent even in the presence of deviations in the initial temperature, solid fuel composition, critical section diameter, etc. Spectrograms of the unwrapped signal were calculated for each sensor to complete the frequency analysis. They are shown in Figure 12. Some peaks in the characteristic frequencies can be expected. The values of the first longitudinal f lon , tangential f tan, and radial f rad modes of chamber sound vibrations, according to Equation (4), are as follows: f lon = from 45 kHz at SRM start to 3.30 kHz at finish f tan = 28.13 kHz f rad = 58.56 kHz A shifting peak, in the range from 3 to 20 kHz, and its harmonics are visible in the spectrograms, and have been caused by the changing longitudinal modes. They have a lower frequency in comparison to tangential and radial modes, so longitudinal modes were the most probable. It is possible to calculate the speed of sound, which is determined by the used fuel. Based on the boundary conditions-the minimum frequency of the longitudinal mode in Figure 12a is 3.3 kHz, and the length of the combustion chamber, which was 109 mm-the following results are found using Equation (4): The spectrograms in Figure 12 show that the peak of longitudinal oscillations only appeared in the interferometer signal when the burning front reached the MZI fixing point. We also analyzed the spectrogram of the audio signal, which was recorded by a microphone during the experiment, and is presented in Figure 13. This plot did not reveal any changing peaks during the burning process. This highlights the advantage provided by the FOS, which was able to detect vibrations that have been generated via sound longitudinal modes. Thus, the fiber MZI worked as a small, light, fire-safe, and easily installed sensor for SRM monitoring.
We plotted the graphs of the ideal longitudinal vibration modes with a known speed of sound and the length of the combustion chamber, calculated according to Equation (4). They are shown in Figure 12b,d,f and are in good agreement with the experimental data. The length changes of the combustion chamber account for two aspects. The first is a meniscus of the burning front, among other factors, caused by a deepening in the solid fuel, as shown in Figure 2. The second is a partial burnout of the plug with the nozzle, from 9 to 5 mm in the center, as shown in Figure 2b. It should be noted that the frequencies of the peaks on the MZI1 and MZI2 spectrograms coincide during their occurrence.
The graphs obtained allow us to conclude that the rate of fuel burnout in the model SRM was almost constant, since the burning front reached the coordinate of the MZI1 after 1 s following initiation.
veal any changing peaks during the burning process. This highlights the advantage provided by the FOS, which was able to detect vibrations that have been generated via sound longitudinal modes. Thus, the fiber MZI worked as a small, light, fire-safe, and easily installed sensor for SRM monitoring.
We plotted the graphs of the ideal longitudinal vibration modes with a known speed of sound and the length of the combustion chamber, calculated according to Equation (4). They are shown in Figure 12b,d,f and are in good agreement with the experimental data. The length changes of the combustion chamber account for two aspects. The first is a meniscus of the burning front, among other factors, caused by a deepening in the solid fuel, as shown in Figure 2. The second is a partial burnout of the plug with the nozzle, from 9 to 5 mm in the center, as shown in Figure 2b. It should be noted that the frequencies of the peaks on the MZI1 and MZI2 spectrograms coincide during their occurrence. (e) (f) The graphs obtained allow us to conclude that the rate of fuel burnout in the model SRM was almost constant, since the burning front reached the coordinate of the MZI1 after 1 s following initiation.
Discussion
A non-invasive diagnostic technique using fiber-optic MZIs as sensors has been developed. This technique makes it possible to determine the characteristics of intra-chamber processes-particularly the burning rate of solid fuel and the length of the combustion chamber-at a given time. The calculations aim to determine when the burning front passes through the MZI fixation points. For each MZI, this can be determined by the derivative growth. Additionally, the resonance frequencies of the acoustic vibration longitudinal modes in the combustion chamber can be determined via the shifting peaks in the spectrogram. As a result, the calculation of the combustion chamber length and the burning rate of solid fuel during the overall SRM worktime can be performed.
For the tested SRM, an uneven combustion of the fuel was detected during the first phase of the work due to deepening occurring at the point at which combustion begins.
Discussion
A non-invasive diagnostic technique using fiber-optic MZIs as sensors has been developed. This technique makes it possible to determine the characteristics of intra-chamber processes-particularly the burning rate of solid fuel and the length of the combustion chamber-at a given time. The calculations aim to determine when the burning front passes through the MZI fixation points. For each MZI, this can be determined by the derivative growth. Additionally, the resonance frequencies of the acoustic vibration longitudinal modes in the combustion chamber can be determined via the shifting peaks in the spectrogram. As a result, the calculation of the combustion chamber length and the burning rate of solid fuel during the overall SRM worktime can be performed.
For the tested SRM, an uneven combustion of the fuel was detected during the first phase of the work due to deepening occurring at the point at which combustion begins. Then, the burnout rate became almost constant; for our experimental conditions, the burnout rate was approximately 0.0193 m/s.
Data Availability Statement:
The data presented in this study are available on request from the corresponding author.
Conflicts of Interest:
The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results. | 7,319.2 | 2021-11-25T00:00:00.000 | [
"Engineering",
"Physics"
] |
Quantum Kinetic Equation for Fluids of Spin-1/2 Fermions
Fluid of spin-1/2 fermions is represented by a complex scalar field and a four-vector field coupled both to the scalar and the Dirac fields. We present the underlying action and show that the resulting equations of motion are identical to the (hydrodynamic) Euler equations in the presence of Coriolis force. As a consequence of the gauge invariances of this action we established the quantum kinetic equation which takes account of noninertial properties of the fluid in the presence of electromagnetic fields. The equations of the field components of Wigner function in Clifford algebra basis are employed to construct new semiclassical covariant kinetic equations of the vector and axial-vector field components for massless as well as massive fermions. Nonrelativistic limit of the chiral kinetic equation is studied and shown that it generates a novel three-dimensional transport theory which does not depend on spatial variables explicitly and possesses a Coriolis force term. We demonstrated that the three-dimensional chiral transport equations are consistent with the chiral anomaly. For massive fermions the three-dimensional kinetic transport theory generated by the new covariant kinetic equations is established in small mass limit. It possesses the Coriolis force and the massless limit can be obtained directly.
Introduction
Wigner function has been originally introduced in deformation quantization of classical mechanics as the substitute of probability density in ordinary quantum mechanics. However, it may take negative values, so that it is considered as the quasiprobability distribution which provides quantum corrections to classical statistical mechanics. It is built with the wave functions satisfying Schroedinger equation [1]. This quantum mechanical formulation can be generalized to field theory by employing quantum fields to construct the Wigner function [2]. For spin-1/2 fermions the constituent fields satisfy the Dirac equation. It turned out that the Wigner function method is a powerful tool in constructing relativistic quantum kinetic theories of spin-1/2 fermions [3][4][5][6].
In heavy-ion collisions quark-gluon plasma is created where the constituent quarks are approximately massless [7,8]. Their properties can be inspected within the chiral kinetic theory (CKT) which leads to an intuitive understanding of chiral magnetic effect (CME) [9][10][11] and chiral separation effect (CSE) [12,13] for the chiral plasma subjected to the external electromagnetic fields. When the chiral plasma is considered as a fluid, the chiral transport equations are also useful to study the chiral vortical effect (CVE) [14] and local polarization effect (LPE) [15][16][17].
The main characteristics of the quarks in quark-gluon plasma are acquired by considering them as massless particles. However, it as an approximation. Therefore, studying the mass corrections is needed. The Wigner function formalism of massive spin-1/2 fermions yields some different covariant kinetic equations [18,19], in contrary to the massless case. This is due to the fact that for massive fermions there are more than one way of eliminating the irrelevant set of field equations derived from the quantum kinetic equation (QKE).
Although four-dimensional (4D) approach has some advantages like being manifestly Lorentz invariant, non-relativistic or equal-time formalism is needed if one would like to solve a transport equation starting from an initial distribution function provided by field equations [20][21][22]. Integrating 4D transport equation over the zeroth-component of momentum is the custom method of constructing the related three-dimensional (3D) transport equation [6,22].
Wigner function of spin-1/2 fermions which are coupled to gauge fields is constructed to be invariant under gauge transformations which leave the Dirac equation intact. The gauge invariant Wigner function satisfies the QKE which depends on the field strength explicitly [4,5]. When one deals with relativistic plasma as a fluid, the vorticity effects should also be taken into account. Although the magnetic and vortical effects are similar, the QKE does not explicitly depend on the rotational properties of fluid in contrary to electromagnetic interactions. Hence, noninertial effects like the Coriolis force are absent. To overcome these difficulties we proposed the modification of QKE by means of enthalpy current either for the massless or massive fermions [23,24]. We obtained the relativistic transport equations and studied the 3D theories which they generate. The chiral formulation was successful in generating a consistent 3D CKT which does not depend explicitly on the position vector and also addresses the noninertial effects like the Coriolis force correctly. The modification of QKE also gives satisfactory results for massive fermions. However, the modified QKE has not been following from an action in contrary to the electromagnetic part. Here, we present an underlying Lagrangian which naturally yields the aforementioned modification of QKE.
3D CKT with the Coriolis force was first presented in [25] by making use of the resemblance between the magnetic field and angular velocity. Then this formulation was derived in [26] from the first principles in a rotating coordinate frame. In [26] it was also shown that spatial coordinate dependence appearing in some CKTs can be removed by an appropriate phase space coordinate transformations.
In [27] CKT in curved spacetime has been derived from the QKE. There it was shown that the noninertial effects and the CVE arise when the observer is in the comoving frame of fluid [27]. In this case our modification terms vanish identically as we have already discussed in detail in [28]. However, in flat spacetime formalism, to generate the noninertial effects one should consider the modified QKE. There is an other issue which should be clarified: Obviously the covariant formalism without the modification leads to CVE correctly [17], so that our modification can be seem to over count the CVE. However, this not the case. Within the modified QKE formalism of [23] in 4D the first order solutions of chiral fields are different from the ordinary ones and they generate the CVE correctly as it was verified explicitly in [23]. When one integrates 4D CKE over the zeroth component of momentum to get the nonrelativistic CKT, the modification terms turn up to be essential in acquiring the CVE correctly. In fact, as it will be explained at the end of section 5, the consistent 3D CKT which results from the 4D formalism would not generate the CVE correctly without the modification terms.
We deal with relativistic plasma as a fluid whose constituents are fermions obeying the Dirac equation. We will introduce the vector field η µ (x) which is minimally coupled to Dirac fields as the electromagnetic vector potential A µ (x). However, it is not a U (1) gauge field. Equations of motion of the new vector potential follow from an action which describes fluid dynamics in terms of self interacting scalar field. Although there are some crucial differences, the action which we consider is mainly introduced in [29]. It is invariant under a gauge transformation which is not the custom U (1) symmetry. Dirac equation coupled to the new vector potential is also invariant under this gauge transformation when the Dirac field is transformed appropriately.
To derive the equation of motion of the Wigner function one employs the equation satisfied by the Dirac field coupled to vector potential and its gauge invariance [5]. As far as the Dirac Lagrangian is considered, the form of the gauge transformations related to η µ (x) and A µ (x) are similar. Hence, we generalize the procedure of [5] to derive the QKE satisfied the Wigner function when both of the η µ (x) and A µ (x) gauge fields are present. Then, we decompose the Wigner function in the Clifford algebra basis and obtain equations satisfied by the fields which are the coefficients of Clifford algebra generators. These equations depend explicitly on the field strengths. Some of these equations can be eliminated and the rest can be used to obtain kinetic transport equations (KTE). In general, until acquiring KTE electromagnetic field strength is defined in terms of vector potentials. Once the KTE are established it is expressed by the electric and magnetic fields which satisfy the Maxwell equations. However, as it will be discussed, for the field strength related to fluid one may proceed in two different ways. The first option is to require that the field strength related to fluid is expressed in terms of vorticity and fluid velocity before deriving KTE. The second option is to establish KTE first and then express the fluid related field strength in terms of vorticity and fluid velocity. When the former method is adopted we find the KTE proposed in [23,24]. The latter method which is similar to the electromagnetic case will be the subject of this work. In this case the massless and massive KTE can be obtained by generalizing the kinetic equations established, respectively, in [30,31] and [19].
We will acquire 3D kinetic equations by integrating the covariant equations over the zeroth component of four-momentum. For chiral fermions we will show that a novel 3D CKT is accomplished in the presence of both external electromagnetic field and fluid vorticity which does not depend explicitly on the spatial coordinates. Moreover, this theory possesses the Coriolis force term and it is consistent with the chiral anomaly. It generates the chiral magnetic and vortical effects correctly. When one deals with massive fermions kinetic equations of vector and axial-vector components of Wigner function depend on the spin four-vector a µ [19]. We provide mass corrections to 3D chiral effects by letting a µ be given with the free Dirac equation for small mass values.
We start with presenting the action which is considered to establish transport equations of a fermionic fluid interacting with electromagnetic fields. In section 3 we focus on the part of action which we claim to govern the dynamical evolution of fermionic fluid. In section 4 we present an outline of the derivation of QKE satisfied by the Wigner function of Dirac fields coupled to two independent vector fields. section 5 is devoted to the study of the kinetic equations of chiral fermions in the external electromagnetic fields by taking into account noninertial properties of the fluid. The relativistic and the 3D chiral transport equations are established. The massive fermions are studied in section 6, where the relativistic equations are integrated over the zeroth component of momentum in the small mass limit by approximating the spin four-vector adequately. Discussions of our results are presented in section 7.
Action
To establish quantum kinetic equation for fermionic fluids in the presence of electromagnetic fields we propose the action (2.1) The first term is the Dirac action, We consider two vector potentials coupled to Dirac fields. One of them is the U (1) gauge field A µ , 3) whose dynamical equations are generated by Q is the electric charge and is the field strength of the U (1) gauge field which is invariant under the gauge transformation The other one is the real four-vector field η α , whose coupling constant is ζ, η α is also coupled to the complex scalar field φ, as follows, In the next section we will clarify how the scalar field φ and the vector-field η α , represent the fluid. We work in Minkowski spacetime with g µν = diag(1, −1, −1, −1).
Fluid
In this section we will describe how the variation of the action with respect to η α and φ fields generate effectively the Euler action of a fluid with Coriolis force, whose constituent particles are Dirac fermions. The scalar field φ is the mean field which represents fluid. The vector field η α is an auxiliary field which will be fixed as in (3.21) below, hence it is considered only at the classical level. Our formulation is mainly inspired from [29], where a covariant action was proposed to describe magnetohydrodynamics as a field theory. However, we are interested in expressing the vector field η α in terms of fluid variables like the enthalpy current. In contrary to [29], in our formulation hydrodynamical quantities are not considered as variables. They will be related to the independent field variables η α , φ by some other means as we will discuss. Therefore we do not need the constraints considered in [29]. Let us first write the complex scalar field in terms of two real fields: Plugging this definition into the action (2.8) yields Observe that it is invariant under the gauge transformation After expressing S φ as in (3.3), variation of S F with respect to ϕ generates the following equation of motion On the other hand variation of S F with respect to the η α yields The mean value of particle number current density operator of the Dirac particles can be expressed in the form Here,ψ,ψ are operators and colons denote normal ordering. The exact form of the distribution function f (x.q) can be obtained from the Wigner function satisfying QKE [2]. One can also describe this system as a fluid. For this purpose let us introduce the fluid four-velocity u α ≡ dx α (τ )/dτ, where τ is the proper time and x α (τ ) is the worldline of a fluid element, so that it satisfies u α u α = 1. (3.8) It can be used to decompose the momentum four-vector as q α = (u · q)u α + q α ⊥ , where q ⊥ · u = 0. Then, due to the momentum integral in (3.7), one gets j α = nu α , (3.9) where n is the particle number density. In principle due to the motion of medium, particle number current density can have an anomalous parts. For example due to rotations there would be a term which depends linearly on the vorticity of the medium. However, we ignore the anomalous contributions to current because we consider only classical fields. In the classical field (mean field) approximation the quantum fieldψ can be considered as a collection of wave-packets constructed by the solutions of Dirac equation. Then, when we deal with fluids composed of spin-1/2 particles, a fluid element which contains a large number of particles but compact enough such that they behave homogeneously, can be taken to coincide with one of the wave-packets. Therefore, in the mean field approximation we can writeψ Now (3.6) can equivalently be written as and (3.5) states that the particle number current density is conserved: In fact, in relativistic fluid dynamics particle number current density without dissipation is given with (3.9) [32].
Equipped with these relations we may now discuss why we consider the action (2.1). The kinetic theory of a neutral plasma can be described in terms of the relativistic scalar electrodynamics with two scalar fields φ 1 and φ 2 which are macroscopic wave functions representing the negative and positive charge carriers [33,34]: If one would like to represent neutral plasma in terms of one scalar field φ instead of φ 1 and φ 2 , it cannot minimally couple to A µ . Then, it should interact with electromagnetic fields in a complicated way [29]. In fact (3.6) represents this interaction because the variation of the action (2.1) with respect to A µ shows that (3.7) is the current which describes how the electromagnetic fields will evolve in plasma: (3.14) By inspecting (3.11) we see that interaction of the fluid with electromagnetic fields is exposed by setting In fact by inserting (3.14) and (3.15) into the action (3.1) we get 16) This shows that scalar field components couple to the derivatives of the electromagnetic fields which would be calculated from the Maxwell equations (3.14) whose charge and current distributions are due to the charged fermions. Hence, the interactions between the scalar field and electromagnetic fields are sensitive how the electric and magnetic fields change spatially and temporally.
Variation of S F with respect to σ leads to where V ′ is the derivative of V with respect to its argument. As in [29] we assume that amplitude of φ varies slowly so that the first term is neglected compared to the second one. Thus we get so that one can express n as a function of the real scalar field σ as Plugging (3.20) back into (3.11) leads to By taking the derivative of (3.18) and employing (3.21), we attain Let us write it as where we introduced and Obviously W αβ vanishes for ordinary functions but in it can also be different from zero [29]: ϕ is the phase of scalar field φ, (3.2), so that it is defined up to some functions Hence, along some θ k (x) = 0, curves, the mixed partial derivatives of ϕ can fail to be continuous. We can ignore the singular curves and set W αβ = 0 without loss of generality. Nevertheless, we can also consider singular cases where the following condition is satisfied, Now by taking the derivative of (3.21), and using it in (3.23) we obtain As far as ν = 0, it yields Acceleration which is defined as the derivative of the velocity u α with respect to the proper time τ, can be calculated by making use of (3.27): Let us compare (3.28) with the (hydrodynamic) Euler equations where ρ is the energy density, P is the pressure. They are related to the specific enthalpy h as nh = ρ + P. Here F α is an external force which can be the gravitational force, electromagnetic force and the "fictitious" force such as the centrifugal or Coriolis force [35]. Except their last terms, (3.28) and (3.29) are identical if From the first law of thermodynamics one knows that for the fluids of only one kind of particle whose total number is conserved and without heat exchange. Suppose that (3.30) is satisfied, then by making use of (3.31) we have It can be expressed as follows, where ξ is a positive constant of integration. We need to express ν analytically in terms of fluid's parameters by respecting the equality (3.34). Let us consider the equation of state P = (γ − 1)ρ, where γ > 1 is the adiabatic index. This relation can be taken as the definition of an ideal fluid [35] and it is consistent with the equation of state resulting from the field equations by choosing V (σ 2 ) adequately [29]. Then, we can write h = γ ρ n . (3.35) with ξ ′ = ξγ. In the mean field approximation we identified a fluid element with a wave packet. Hence the proper energy density can be parametrized in terms of the momentum p µ of the wave packet center as From (3.21) by setting W αβ = 0 or employing the condition (3.25), we get , is the relativistic generalization of Coriolis force per particle. We consider vanishing linear acceleration so that the fluid velocity can be taken to satisfy Fluid vorticity four-vector is defined as where is the kinematic vorticity tensor. Therefore we get In the frame u α = (1, 0), ω α = (0, ω), it becomes K β = (0, K) with Hence, we conclude that the last terms of (3.28) and (3.29) are identical where is a relativistic extension of the Coriolis force in a rotating coordinate frame. From (3.39) and (3.38) we have On the other hand, due to the vanishing of linear acceleration, (3.40), we observe that Hence, (3.39) is satisfied for an arbitrary constant κ, which can be even zero. By introducing w µν can be expressed as It is worth noting that ξ, κ are arbitrary constants and w µν is the circulation (vorticity) tensor for κ = 2γ, and ξ = 1, i.e. ξ ′ = γ, [35]. Therefore, we come to the conclusion that the scalar field φ and the vector field η α represent the fluid composed of the Dirac particles. Moreover, the field strength of η α is given by (3.46) when the equations of motion resulting from the variation of S F with respect to φ and η α are satisfied.
Quantum Kinetic Equation
Let us return to the initial action (2.1) without imposing the equations of motion derived from (2.4) and (3.1). Now, we would like to examine the action of Dirac field coupled to the vector fields, It generates the Dirac equation which is invariant under the gauge transformations (2.6) and (3.4), when the spinor field transforms as The Wigner operator is defined bŷ Here, ψ(x) andψ(x) are operators and ⊗ represents tensor product. The Wigner function is defined as the ensemble average of the normal ordered Wigner operator: W (x, p) = :Ŵ (x, p) : .
By making use of the Dirac equation (4.2), one can show that the Wigner operator, (4.7), satisfies where ∂ µ p ≡ ∂/∂p µ . F µν and w µν are defined by (2.5) and (3.24). We consider the mean field approximation, so that the field strengths F µν , w µν are c-valued fields. Following [5] we write f (x + sy − y/2)] = e (s−1/2)y·∂ f (x) and employ the relation to express the right-hand-side of (4.8) as One expands the exponential as power series and perform the s integration. Then by taking the ensemble average of (4.8), the quantum kinetic equation satisfied by the Wigner function is established as j 0 (x) and j 1 (x) are spherical Bessel functions in ∆ ≡ 2 ∂ p · ∂ x . The space-time derivative ∂ µ contained in ∆ acts on [QF µν + ζw µν ] , but not on the Wigner function. In contrary ∂ ν p acts on the Wigner function, but not on [QF µν + ζw µν ] .
The Wigner function can be decomposed by means of the 16 generators of the Clifford algebra where the coefficients C ≡ {F, P, V µ , A µ , S µν }, respectively, are the scalar, pseudoscalar, vector, axial-vector, and antisymmetric tensor fields. We expand them in powers of Planck constant and keep the leading and next to the leading order terms in . Hence in (4.9) one sets π µ = p µ and substitutes D µ with Plugging (4.10) into (4.9), yields complex valued equations whose real parts are 2∇ · A + mP = 0, (4.16) and the imaginary parts are ∇ · V = 0, (4.17) p · A = 0, (4.18) To derive QKE (4.9), we started from the Dirac equation (4.1). Then in getting (4.8) we had to introduce F µν and w µν which are defined in terms of the gauge fields as in (2.5) and (3.24). The equations of motions following from the action (2.4) give the Maxwell equations when one can expresses the field strength in terms of the electromagnetic fields E µ , B µ by where u µ is the fluid 4-velocity. Similarly, when the equations of motion discussed in section 3 are imposed, one deals with the Euler equations (3.29), where w µν is expressed in terms of vorticity and energy per particle as in (3.46). However, when the equations of motion following from (2.4) and (3.1) should be imposed? We have two options: i) Obtain the kinetic equations which the field components of Wigner function, C, satisfy and then impose the equations of motion. ii) Impose the equations of motion from the beginning and then derive the kinetic equations satisfied by the fields C. Our previous works [23,24] are consistent with the latter option for η µ . Although we kept F µν as in (2.5), w µν was expressed in terms of enthalpy current, (3.44), and then derived the kinetic equations satisfied by the fields C. Here, we adopt the former option for both of them: We use the definition of w µν in terms of the η µ fields, (3.24), to establish the kinetic equations and then apply the equations of motion yielding (3.44). Similarly, F µν is given by (2.5) until the KTE are derived. Afterwards, we express it in terms of the electromagnetic fields, (4.22). Once we choose this method the semiclassical kinetic equations can directly be read from the known ones [19,30,31,36], by substituting QF µν with QF µν + ζw µν , as it will discussed in the subsequent sections.
Chiral Kinetic Equations
For vanishing mass the equations of A µ and V µ decouple from the rest. They are given by where χ = 1, and χ = −1, correspond to the right-handed and the left-handed fermions. They need to satisfy The chiral kinetic equation which results from (5.1)-(5.3) can be acquired by generalizing the formalism given in [30,31,36]. First, one can verify that the solution of (5.1) and (5.3) is is the distribution function. n µ is an arbitrary vector satisfying n 2 = 1 and S µν (n) = 1 2n · p ǫ µνρσ p ρ n σ .
Then, by inserting (5.4) into (5.2) one acquires the chiral kinetic equation: whereF µν = 1 2 ǫ µναρ F αρ andw µν = 1 2 ǫ µναρ w αρ , are the dual field strengths. Until now we have w µν = ∂ µ η ν − ∂ ν η µ , because η µ was off-shell. Now, we impose the equations of motion (3.11)-(3.17), thus w µν is given by (3.46). In the rest frame of massive fermions energy per particle is m. Therefore, in the massless limit we set κ = 0, and write where k = −ζκ is an arbitrary constant which will be fixed. Now, (5.5) is the chiral kinetic equation where the vorticity and electromagnetic tensors are treated on the same footing.
To establish a 3D CKT we would like to integrate (5.9) over p 0 . To perform this integration we decompose the distribution function into particle s = 1 and antiparticle s = −1 parts, Moreover, we choose the frame u α = (1, 0) and ω α = (0, ω), (5.12) where the delta function yields the dispersion relations
Let us also introduce the canonical velocity
(5.14) As we show in appendix A, integrating (5.9) over p 0 leads to the transport equation We ignore the O(ω 2 ) terms. The chiral particle (antiparticle) number and current densities are defined by To accomplish the continuity equation satisfied by the 4-current density j χµ s ≡ (n χ s , j χ s ), let us calculate Observe that (5.16) and (5.17) do not depend on time and spatial coordinates explicitly, so that we have By employing the transport equation (5.15) one gets The derivative of (5.18) leads to By substituting the first term in (5.22) with (5.23) one finally attains On the other hand the last term of (5.21) vanishes because it is a total derivative, thus the continuity equation is deduced: Let us introduce the vector and axial-vector currents where j R = s=±1 j 1 s and j L = s=±1 j −1 s . Let us choose the equilibrium distribution function as f eq,s Here µ χ = µ + χµ 5 , where µ, µ 5 are the total and axial chemical potentials. By inspecting (5.17) one observes that the currents are linear in the magnetic field and vorticity: The coefficients of the magnetic field are calculated as Thus the magnetic terms in (5.28) and (5.29), respectively, generate the chiral magnetic and chiral separation effects. The vorticity terms in (5.28) and (5.29), respectively, generate the chiral vortical and local polarization effects correctly for ξ = µµ 5 π 2 2 , ξ 5 = However, the coefficients ξ and ξ 5 depend on k. One can show that (5.30) results as far as the condition 2 3 is satisfied. This yields k = 1. This value of k is in accord with the formalism considered in [23]. We do not deal with the equilibrium distribution function for a rotating fluid because in constructing the CKE (5.7) we have not considered rotation of the reference frame. However, in the absence of modification terms one should work with the equilibrium distribution function of a rotating fluid in the comoving frame of the fluid [36,37] f eq,s is fulfilled, which yields k = 1/2. Obviously, for the original CKE where k = 0, the condition (5.34) cannot be satisfied. Therefore, we conclude that without the modification the 3D formalism obtained from (5.7) does not generate the CVE correctly. This result is pertinent to the transport equation in Minkowski space with the condition n µ = u µ , where u µ and ω µ are given as in (5.12). Obviously, one can choose either n µ or u µ in a different manner, e.g.
. For each choice one should derive the resulting 3D transport equations by integrating (5.5) over p 0 .
As it was mentioned in Introduction there also exists a curved spacetime formulation of chiral kinetic equation [27] where the Coriolis force and CVE are generated correctly without a need for modification. However, this formalism leads to a 3D kinetic theory which depends on x explicitly [28], in contrary to the 3D kinetic theories obtained here (5.16)-(5.18) or as it has been derived in [23].
Kinetic Equations of Massive Fermions
The equations (4.12)-(4.21) which should be satisfied by the components of Wigner function in Clifford algebra basis are reducible: The field equations (4.12), (4.15) and (4.16) can be employed to express the fields F, P, S µν in terms of the vector and axial-vector fields, V µ , A µ . Following the procedure given in [22] and [19] the rest of field equations (4.12)-(4.21), can be shown to yield∇ · V = 0, (6.1) where at most O( ) terms are taken into account. We can derive the semiclassical kinetic equations resulting from (6.1)-(6.6) by substituting QF µν with QF µν + ζw µν in the formalism which has been given in [19] for ζ = 0. First one solves (6.3), (6.4) for V µ up to -order, then uses it in (6.1) and gets the kinetic equation of the vector field: Here f V , f A are vector and axial distribution functions.
S µν a(n) = 1 2n · p ǫ µνρσ a ρ n σ (6.8) is the spin tensor and a µ is the spin four-vector which is defined to satisfy the constraint To derive the other kinetic equations which are needed to determine the dynamical degrees of freedom f V , f A , a µ , one solves (6.2) and (6.5) for A µ up to O( ) with the help of Wigner function of the quantized free Dirac fields [19]. Plugging this solution into (6.6) leads to where S µν m(n) = 1 2(n · p + m) ǫ µνρσ p ρ n σ . (6.11) Once we get these kinetic equations, we impose the equations of motion resulting from (2.4) and (3.1). Thus, from now on we deal with F µν expressed in terms of the electromagnetic fields E µ , B µ as in (4.22) and the field strength w µν expressed in terms of the vorticity ω µ and the energy per particle u · p, as in (3.46). To keep the discussion general let us introduce ζw µν ≡ Λw µν C + 2εu · p Ω µν . (6.12) How to fix the constants ε and Λ will be discussed later. By making use of (5.10) and the Schouten identity, (6.12) can be written equivalently as The kinetic equations (6.7) and (6.10) form the 4D collisionless kinetic theory of massive fermions in the presence electromagnetic fields and vorticity. We would like to derive the nonrelativistic kinetic theory arising from this covariant formulation. For this purpose let us extract the scalar kinetic equation stemming from (6.10) by projecting it on n µ : We will show that (6.7) and (6.14) can be combined to derive scalar kinetic equations which generate correct dispersion relations of Dirac fermions coupled to electromagnetic fields in a frame rotating with the angular velocity ω.
3D Massive kinetic theory
To establish the 3D kinetic theory arising from the relativistic kinetic equations (6.7) and (6.14), one should determine the form of a µ , consistent with the kinetic equations and the constraint (6.9). However we are only interested in attaining small mass corrections to the chiral effects. Hence, let us deal with a µ derived from the Wigner function of the free Dirac fields [19]. In this case it is defined to satisfy where S µ is the spin four-vector of the Dirac wave-function and a ⊥ · n = p ⊥ · n = 0. Thus, it can be expressed as a µ = mn µ + p µ − mS µ .
S · p = p · n + m, so that (6.9) is fulfilled. Moreover, let S µ be given as in the massless case: Therefore, we set Because of dealing with a µ at the zeroth-order in electromagnetic fields and vorticity, we need to keep the terms which are at most linear in F µν and w µν in the kinetic equations (6.7) and (6.14). For simplicity, we consider the fields satisfying ∂ µ F νσ = 0 and ∂ µ ω ν = 0.
To establish the 3D kinetic theory by integrating over p 0 , we have to work in the comoving frame by setting n µ = u µ . (6.17) Let us accomplish the kinetic equations following from (6.7) and (6.14) in the comoving frame (6.17), by keeping at most the terms linear in E µ , B µ and ω µ . The kinetic equation of the vector field (6.7) reads with the spin tensor The left-hand side of (6.14) can be written with an overall (u · p + m) factor which is obviously nonvanishing. Hence (6.14) yields the kinetic equation By expanding the denominators for small mass with m/u·p ≪ 1, and ignoring the m 2 /(u·p) 2 and higher order terms, (6.19) can be written as In the small mass limit we have Then, by summing and subtracting (6.18) and (6.20) we acquire the kinetic equations Here f χ ≡ f R/L = (f V +χf A )/2 are the right-handed and left-handed distribution functions. The right-hand side of (6.21) vanishes for both m = 0 and = 0, but its left-hand side is non-zero whether = 0 or m = 0. Moreover, the distribution functions appearing on the left-and the right-hand sides possess opposite chirality. Hence we can consider (6.21) as the transport equation of f χ which shows up on the left-hand side where its right-hand side is due to presence of the opposite chirality distribution function f −χ , for massive fermions.
To derive the nonrelativistic kinetic equations by integrating (6.21) over p 0 , we consider distribution functions composed as in (5.11) and choose the frame u µ = (1, 0), ω µ = (0, ω). In this frame the Dirac delta function yields the dispersion relations where E p = p 2 − m 2 and b ≡ p/2E 3 p . Calculation of the p 0 integral yields √ η s The right-hand side of (6.24) which vanishes for m = 0 can be considered as the correction terms due to the presence of opposite chirality fermions in the massive case. Terms which vanish for m = 0 can also be interpreted as collision terms due to interaction of electromagnetic fields and vorticity with the spin of massive fermions [38]. Let us fix the values of ε and Λ. Comparing the dispersion relations (6.23) with the Hamiltonian of Dirac particles coupled to the magnetic field in rotating coordinates in the helicity basis [24,39], one observes that ε = 1/2. Similar energy relations were also obtained for chiral particles in [27,40]. Now, by inspecting (6.27) we see that Λ = 1 for reproducing the Coriolis force correctly. This choice is also consistent with the massless case considered in the section 5.
Inserting the spatial velocity (6.26) into the following definition where f eq,s χ is given in (5.27) by substituting E χ s with (6.23), we write the particle number axial-vector and vector current densities as They can be decomposed as follows where f 0 A,V are defined by For the sake of simplicity let us deal with µ R = µ L = µ. Whileσ B V vanishes under this condition,σ ω V isσ at zero temperature. By performing the integrals at zero temperature, the coefficients of axial-vector current density are calculated as A is in accord with the one derived by means of Kubo formula [41][42][43]. For ε = 1/2, Λ = 1 we haveσ Note that when the massless limit is considered one should also set ε = 0. For instance for small m, by setting ε = 0, one obtains This is the correction obtained in [43,44] in the µ = 0 limit. As we have already mentioned there is not a unique method of defining kinetic transport equations for massive fermions beginning with (4.12)-(4.21). Hence there is no consensus in obtaining the mass corrections to the chiral magnetic and vortical effects. Here we established the massive kinetic equation from the covariant formulation which directly generates the chiral transport equation when one sets a µ = p µ and m = 0. In fact we have chosen the spin four-vector as in (6.16) which is consistent with the massless limit. Therefore the corrections to the chiral effects are apparent in contrary to the other formalisms which are suitable to discuss large mass limit [18,23].
The massless limit can also be discussed starting from (6.7) and (6.14). For m = 0 we need to set ε = 0 in (6.12) and fix the spin four-vector as a µ = p µ . One can then observe that (6.14) yields a scalar kinetic equation up to an overall p µ factor. A similar scalar kinetic equation follows from (6.7). By adding and subtracting these scalar kinetic equations one obtains the chiral kinetic equation (5.7) for n µ = u µ and Λ = k. Therefore the 3D chiral theory coincides with the one obtained in section 5, in particular particle number current density satisfies the anomalous divergence as in (5.25).
Conclusions
We demonstrated that the scalar field φ and the vector-field η µ represent the fermionic fluid in the presence of the Coriolis force due to the nonvanishing vorticity. This is achieved by showing that the equations of motion acquired by variation of the action (3.1) with respect to φ and η α are equivalent to relativistic Euler equations of a fluid with the Coriolis force. Moreover, this formalism provided the field strength of the vector field η α in terms of the specific enthalpy and fluid vorticity.
Then we considered the action of Dirac spinors coupled to the vector fields A µ and η µ . By virtue of its gauge invariances we derived the QKE satisfied by the Wigner function, (4.9), by generalizing the formalism of [5]. In fact, one of the main results accomplished in this work is to show that when one deals with the fluids the original QKE [5] should be modified adequately. This modification has already been introduced in [23,24] in an ad hoc manner. Here we obtained it in a systematic way from an underlying action.
The Clifford algebra generators are employed to decompose the Wigner function in terms of field components and the semiclassical equations of the component fields which follow from QKE are presented. Then to derive KTE one can proceed in two different ways depending on when to impose the on-shell conditions of fields representing the fluid. In [23,24] we have derived the semiclassical kinetic equations by imposing the on-shell conditions from the start. In contrary here we first acquired the semiclassical transport equations and then let the fields η α , φ be on-shell, so that w µν is expressed in terms of the vorticity and fluid 4-velocity. This approach furnishes novel kinetic transport equations for either massless or massive fermions.
As usual the equations satisfied by vector and axial-vector decouple from the other fields when one considers massless fermions. The kinetic equation in the presence of unique gauge field is well known [30,31,36]. We generalize it and obtain the semiclassical chiral kinetic equation where the electromagnetic fields and the vorticity are treated on the same footing. By integrating the semiclassical relativistic kinetic equation over p 0 , we established a 3D CKT which does not depend on the spatial coordinates explicitly. It is consistent with chiral anomaly and takes into account the non inertial properties of rotating reference frame. Moreover, the chiral magnetic and vorticity effects are correctly generated.
Transport equations of massive fermions are also studied. The semiclassical kinetic equations of the vector and axial-vector fields are derived by extending the formalism given in [19]. The related 3D kinetic transport equation is obtained by letting the spin 4-vector be adequate to discuss the small mass limit. We showed that the Coriolis force and the dispersion relation are correctly generated. We obtained the particle number current density in terms of the equilibrium distribution function and calculated it at zero temperature. The similarities and differences with the other approaches are discussed.
For massive fermions we obtained the 3D transport equations by adopting the definition of spin vector as in (6.16) and keeping only the linear terms in the electromagnetic fields and vorticity. For having a better understanding of the mass corrections to the chiral effects, one needs to find a solution of the spin vector depending on the electromagnetic fields and vorticity.
Although we considered the collisionless case, kinetic transport equations are mainly needed in the presence of collisions. They can be introduced in the current formalism by generalizing the methods given in [30] and [45]. Obviously collisions can also be studied within the 3D kinetic transport theories.
Zilch vortical effect [46] has been recently studied within the kinetic theories in [47,48]. In [47] it was shown that the zilch current in a rotating system can equivalently be derived in chiral kinetic theory by employing chiral current. It would be interesting to inspect if this construction of zilch vortical effect can be studied by means of our approach for example by modifying the chiral current. This may provide an alternative way of studying the rotation of reference frame for photonic media. | 9,846.2 | 2021-06-24T00:00:00.000 | [
"Physics"
] |
2024 Cosmological fluids with boundary term couplings
Cosmological models can be studied effectively using dynamical systems techniques. Starting from Brown’s formulation of the variational principle for relativistic fluids, we introduce new types of couplings involving a perfect fluid, a scalar field, and boundary terms. We describe three different coupling models, one of which turns out to be particularly relevant for cosmology. Its behaviour is similar to that of models in which dark matter decays into dark energy. In particular, for a constant coupling, the model mimics well-known dynamical dark energy models while the non-constant couplings offer a rich dynamical structure, unseen before. We are able to achieve this richness whilst working in a two-dimensional phase space. This is a significant advantage which allows us to provide a clear physical interpretation of the key features and draw analogies with previously studied models.
Introduction
Cosmology, the scientific study of the universe as a whole, has undergone remarkable advances in recent decades and General Relativity (GR) provides a good model to describe cosmological gravitational phenomena [1,2,3,4].On the other hand, open questions in cosmology remain, foremost among which are the dark energy and dark matter problems.The nature of dark energy, which is responsible for driving the universe's late-time accelerated expansion, is not well understood, and it often assumed to be a cosmological constant.Since the first observational evidence of an accelerated expansion [5,6] of the universe, a plethora of cosmological models to explain dark energy has emerged, for a review see [7].
The addition of a positive cosmological constant Λ to the Einstein field equations, originally introduced by Einstein [8] for his static universe, is one of the most straightforward candidates for dark energy.This paves the way for the Λ Cold Dark Matter (ΛCDM) model.However, the ΛCDM model fails to explain why the inferred value of Λ is so small compared to the vacuum energy density expected from particle physics [9].It is also unclear why its value is comparable to the matter density today.This constitutes the so-called coincidence problem [10,11].
One way to begin to address this issue is to allow for a dynamical cosmological constant [7], that is, to introduce some dynamical field able to reproduce the late-time acceleration behaviour and mimic the properties of the cosmological constant.The simplest such model is a canonical scalar field φ with flat potential V (φ), which drives the accelerated expansion of the universe.Any model of this type is referred to as quintessence [12].Scalar fields play a major role in modern cosmology as they are also able to drive inflation, the early-time epoch of accelerated expansion [13,14,15].Scalar field models have also been used as candidates for dark matter models, see [16].We are primarily interested in scalar fields as models to drive a period of accelerated expansion, at both early and late times of the universe's evolution.
Another approach is to consider dark energy as evidence for the incompleteness of GR and, hence, seek extensions or modifications of GR [17,18,19].Several models to describe the dark energy interaction with dark matter have been proposed [20,21,22,23,24,25,26,27] while some authors, e.g.[28], have emphasised a strong distinction between modified theories of gravity and dark energy models.
From a theoretical point of view, in GR, one usually restricts the Lagrangian to a linear function of the Ricci scalar, minimally coupled with matter.However, there is no reason, a priori, to assume such a restriction.So, one can modify the gravitational part of the action to allow non-linear corrections to the Lagrangian [29,30], this is the general approach followed by f (R)-theories of gravity [31,32].Some other extensions of GR increase the number of spacetime dimensions or introduce non-minimal matter couplings to boundary and topological terms [33,34,35,36,37,38,39,40,41,42].These are terms in the Lagrangian that describe how matter couples to geometrical quantities.Non-minimally coupled terms involving curvature vanish in the limit of special relativity.
We will follow the approach suggested in [43,44], and, therefore, build models of quintessence interacting with dark matter.This involves introducing couplings at the level of the action which characterise both quintessence and dark matter [45,46].In particular, we extend the Lagrangian formulation by Brown [47] which describes a perfect fluid [48].Motivated by [43,44], we introduce new couplings containing a boundary term and a pseudovector related to the boundary term.
In [40,49], a dynamical system analysis where teleparallel quintessence is non-minimally coupled to a boundary term is presented.In the same spirit, we study the background cosmology within this framework and apply dynamical systems tools to investigate the dynamics of the different models.Our ultimate goal is to examine the behaviour of these dark energy models.Dynamical systems theory has emerged as a vital tool in cosmology [50], and has been employed successfully to study modified theories of gravity in the cosmological context [51,52,53].
The paper is organised as follows.In Section 2, we present the Lagrangian formulation of our models, and, in Section 3, we focus on the cosmological field equations and introduce the cosmological variables which will be used for our analysis.Section 4 contains our analysis of the dynamical systems for a constant interaction term, and we highlight the analogies with previous models [54].In Section 5, we present the rich and novel dynamical structure in the case of a non-constant interaction term, which, for some choice of the parameters, features both early-time and late-time accelerated expansion.In Section 6, we discuss our results and suggest potential directions for future investigations.
Notation and conventions.Unless otherwise specified, we employ standard relativistic notation throughout.The signature of the metric tensor g µν is assumed to be (−, +, +, +), Greek indices are space-time indices taking values in {0, 1, 2, 3}.The coupling constant appearing in the Einstein field equations is denoted by κ = 8πG/c 4 , where c is the speed of light and G the Newton's gravitational constant.We use natural units, with c = 1 and G = 1.A dot denotes differentiation with respect to cosmological time, a prime denotes the derivative with respect to the argument, or in the case of the dynamical system equations a prime denotes a derivative with respect to the logarithm of the scale factor log(a).
2 Lagrangian formulation and field equations
Gravitational and fluid action
It is well known that a total derivative term can be isolated from the Ricci scalar, yielding the Gamma squared action.This action also gives rise to the Einstein field equations when variations with respect to the metric are considered.However, the underlying Lagrangian is no longer a coordinate scalar as it differs from a true scalar by the total derivative term.We prepend the word 'pseudo' to highlight quantities which appear to be scalars or tensors but are not invariant under general coordinate transformations.As shown in [42], this allows one to write the Einstein-Hilbert action as The bulk term G is defined as and the boundary term B is given by where we have introduced the boundary pseudovector B σ given by We note that G and B are pseudoscalars.By construction, the bulk term G is quadratic in the Christoffel symbols and hence the action is called the Gamma squared action or, sometimes, the Einstein action, to distinguish it from the Einstein-Hilbert action.Recent progress was made in [42,55] on constructing modified theories of gravity based on this decomposition.These models can be linked naturally to a variety of other modified gravity models, either within the context of GR or in the metric-affine framework.The Christoffel symbols are usually interpreted as the gravitational field strengths.We can motivate this by recalling that they contain the first partial derivatives of the metric, which represent the gravitational potentials.The bulk term is thus quadratic in the field strengths, similar to other field theories, like Yang-Mills theories or elasticity theory.This analogy provides the primary motivation for splitting the Ricci scalar as in (2.1).This split naturally yields a boundary term which could be coupled to other fields present in the model.Couplings of this type are interesting as they are purely geometrical and thus have no direct links with classical physics.This is similar to Brans-Dicke theories, where a scalar field is coupled to the curvature scalar.By isolating the bulk and boundary terms, we can therefore consider more intricate couplings involving those two parts, which make up the curvature scalar.
Matter is introduced into the theory by considering a total action of the form where S matter is the matter action.This gives rise to the energy-momentum tensor Generally, a matter action would depend on matter fields, and variations with respect to those matter fields yield the equations of motion of the matter component.As we will see in the following, this is non-trivial if one wishes to model relativistic fluids using the variational approach.
Brown [47] introduced a Lagrangian formalism for relativistic perfect fluids based on the Lagrangian (density) given by where • n is the particle number density • s is entropy density per particle • ρ(n, s) is the energy density of the matter fluid, a function of n and s • J µ is the vector-density particle-number flux, which is related to n by where U µ is the fluid's 4-velocity satisfying U µ U µ = −1 • ϕ, θ, and β A are all Lagrange multipliers with A taking values in {1, 2, 3}, and the components α A are the Lagrangian coordinates of the fluid.
The independent dynamical variables of the Lagrangian (2.8) are g µν , J µ , s, ϕ, θ, β A , and α A .We note that, in this approach, the pressure of the fluid p is defined as which is consistent with the first law of thermodynamics.
Total action and interaction terms
We can now set up the total action which contains gravity, a fluid, a scalar field φ, and an interaction term.This means where L φ is the scalar field Lagrangian (density) given by with given scalar field potential V .The Lagrangian (density) L int is an interaction coupling term, which allows us to couple the fluid to the scalar field.Due to the presence of the various independent variables in Brown's approach, one can propose some types of coupling terms which do not exist in other settings.Moreover, such terms have no natural special relativistic analogue, making this potentially interesting in the context of cosmology.In previous work [43,44], one of the authors proposed interaction terms of the form f (n, s, φ) and f (n, s, φ)J µ ∂ µ φ.These models gave rise to some unexpected dynamics.In particular, as we wish to take into account boundary terms, we identified the following terms as the suitable possibilities Note that B µ was defined in (2.4).Depending on the specific interaction term chosen, one should also note that the physical dimensions of f differ in the different couplings.
Coupling (b) is very restrictive in the context of cosmology.We find that the consistency of the cosmological equations implies that f is proportional to n, thereby eliminating the scalar field from the coupling.Consequently, we find equations which largely coincide with the standard cosmological equations, and the model does not exhibit novel behaviour.
For the remainder of this paper, we will consider the interaction term (c) and L int will denote this interaction Lagrangian.This term was found to have behaviour relevant to cosmology, and gave rise to manageable cosmological equations.In principle, our analysis can be repeated for term (a), and potentially for more complicated terms.For example, f could contain an explicit dependence on G, or the Ricci scalar, or there could be higher order couplings containing terms like (B µ J µ )(B ν ∂ ν φ), etc.
Variations and field equations
We begin with the variations of action (2.11) with respect to the the fields ϕ, θ, β A , and the Lagrangian coordinates α A , respectively.This yields δϕ : J µ ,µ = 0 , (2.13) δθ : (sJ µ ) ,µ = 0 , (2.14) ) These equations are independent of the gravitational action and are also independent of the interaction term.Next, variations of (2.11) with respect to the entropy density, s, give where the final term depends on the choice of f .Variations with respect to J µ yield Again, we have one term which depends on the coupling.Variations with respect to the scalar field φ yield a modified Klein Gordon equation where := ∇ µ ∇ µ .
Finally, variations with respect to the metric tensor yield the Einstein field equations where G µν is the Einstein tensor and Both are the standard forms of the energy-momentum tensors of a perfect fluid and a scalar field, respectively.The energy-momentum tensor related to the interaction term is more complicated and is given by The second line of this tensor appears due to variations of the boundary pseudovector with respect to the metric.This requires integration by parts and thus leads to the second derivative terms of the scalar field.We note that one could rewrite the partial derivatives of the metric determinants using the Christoffel symbols.However, for the purposes of this work, this would not introduce additional insights.
We motivated the introduction of L int as a new interaction term which could model an interaction between the fluid and the scalar field.However, one can adopt a different interpretation, namely to view (2.23) as an independent fluid of unusual form.
Cosmological field equations
In this section, we provide a brief overview of the necessary background material required to study the cosmological field equations of our coupled models.We do this via a dynamical systems formulation, which has proved to be a powerful tool when studying cosmological equations.
In line with current observational evidence [2,56,57], let us begin with the homogeneous, isotropic, and spatially flat Friedmann-Lemaître-Robertson-Walker (FLRW) line element where a(t) is the scale factor and N (t) is the lapse function.For all models under consideration, we will be able to set N = 1, which simplifies the cosmological equations further.In this case, t is cosmological time.In this cosmological setting, (2.20) yields the cosmological Einstein field equations given by and (2.19) leads to the modified Klein-Gordon (KG) equation Here the dot denotes differentiation with respect to cosmological time, and we remark that ρ and p are the energy-momentum density and pressure of the fluid, respectively.
A direct, but lengthy, calculation verifies that the three equations (3.2)-(3.4)imply the fluid's energy-momentum conservation equation ρ + 3H(ρ + p) = 0.This is a non-trivial result which is, perhaps, unexpected given that the coupling contains an unspecified function.Let us also note that the only dependence on the scale factor a(t) in the field equations is via the Hubble function H and its derivative.These equations feature both first and second derivatives of the scalar field, φ.However, following [54], one can introduce a new variable which depends on the first derivative of the scalar field, leading to field equations which are first order.In short, this is the key idea behind the dynamical systems formulation.
If we assume that the potential is non-negative, we can introduce the well-known dimensionless variables, first proposed in [54], We restrict to the case that H > 0, i.e. that the universe is expanding (choosing H < 0 would correspond to a contracting universe).It follows that the variables y and σ are non-negative.
In line with previous studies, we assume V has the exponential form where V 0 > 0 is a constant and λ ≥ 0 is a dimensionless parameter.We note that this form for V is invertible, which will allow us to view φ as a function of V .This potential is most convenient as the exponential form allows one to close the autonomous system of equations without the introduction of an additional variable.
When the FLRW metric is considered in (2.13) and (2.14), one immediately finds that the entropy density s = s 0 is a constant.Consequently, the coupling function is of the form f (n, φ) only.Moreover, (2.13) and (2.14) also imply that the particle number density is n = n 0 a −3 , where n 0 is a constant, which is expected.
Going back to Brown's formulation (2.8), we have that the energy density is a function of n, since the fluid's entropy s is constant, thus ρ = ρ(n).On the other hand, in standard cosmology, it is customary to assume a linear equation of state of the form p = wρ.We will now demonstrate that, given the definition of pressure in (2.10), this is equivalent to the assumption that the density is a power of the particle number density.To begin with, let us consider ρ = n w+1 , for some w, which implies For the matter dominated case, w = 0, this gives that p = 0. On the other hand, integrating (2.10) with the assumption that p = wρ implies ρ = n w+1 .Dividing equation (3.2) by 3H 2 and using the variables (3.5), one obtains We note that f must be chosen to have the same dimensions as κ −1/2 to ensure that this equation is consistent.As it is derived from the Friedmann constraint equation, we will generally refer to it as the constraint equation.This is motivated by the fact that it is an algebraic relation between all the variables, which implies that the variables are not all independent.We finish this Section by noting that for f (n, φ) = 0, we retrieve the model studied in [54], which we can view as our baseline model.When interpreting our results, we draw analogies and highlight differences with this baseline model.In that work, the constraint equation (3.8) is solved for the matter variable σ which is then eliminated from the other equations, reducing the system to two differential equations and we follow the same approach here.
Constant interaction
To begin our study, we consider perhaps the simplest non-trivial model, where the coupling function is a constant for some constant k.This model shares some similarities with [54] and is an ideal prelude to the study of more complicated models.
General properties and dynamical systems formulation
Let us start with the Klein-Gordon equation (3.4), which simplifies to where we introduce the quantity Q to match previous work on dark sector couplings [45].The energy density and pressure of the scalar field are given by ρ φ = φ2 /2 + V and p φ = φ2 /2 − V , respectively.This allows us to re-write equation (4.2) in the well-known form hence Q can be re-expressed as where q = −1 − Ḣ/H 2 is the standard deceleration parameter.It is well known from dark sector coupling models [45,46] that Q > 0 means an energy transfer from dark matter to dark energy and Q < 0 a transfer in the opposite direction.
For k > 0, equation (4.4) implies that an epoch of accelerated expansion, q < 2, gives a positive coupling, leading to energy going into the scalar field.In turn, this leads to an epoch of further acceleration and can be seen as a self-reinforcing effect.The above argument is reversed for f < 0 (i.e.k < 0).Given that the late-time universe is dark energy dominated while the early universe contains considerably more dark matter than dark energy, it is reasonable to consider f > 0 (i.e.k > 0) and it will turn out that such models indeed evolve into epochs of late-time accelerated expansion.
Next, we consider a fluid with equation of state p = wρ, which, as discussed at the end of Section 3, is equivalent to setting ρ = n 1+w .The constraint equation (3.8) then reads The quantity σ 2 is the relative energy density of matter, sometimes denoted by Ω m when discussing explicit cosmological models.For a scalar field, it is helpful to introduce the equation of state Here, w φ ∈ [−1, 1] and we get w φ = −1 when φ = 0, as is expected for dark energy.The energy density of the scalar field is given by Ω Hence, (4.5) can also be written as At this point, it is clear that one can introduce improved variables by completing the square of the x-term in (4.5).Namely, we write and now divide by the new right-hand side so that we arrive at where These variables will prove particularly useful for our subsequent qualitative analysis.Using equations (3.2)-(3.4),one can obtain the acceleration equation which can be integrated to find a(t) at any given fixed point (X 0 , Y 0 ).The right-hand side, at a fixed point, is constant.If this constant is non-zero, it is straightforward to show that the scale factor evolves as a power law in cosmological time, that is, a ∝ (t − t 0 ) µ , where µ is that power.We therefore have that µ is given by Here t 0 is an integration constant.When the right-hand side of (4.12) vanishes at some fixed point, the scale factor a(t) evolves exponentially.This corresponds to H being constant at this point, that is, a universe undergoing a de Sitter expansion.
It can be useful to define the total energy density and total pressure of the cosmological model where we set ρ int = −kH 6/κ + φ and p int = k φ/ √ 6κ, as suggested by (3.2) and (3.3).This naturally leads to the effective equation of state parameter w = p/ ρ.For power law models, this effective equation of state parameter is directly related to the power µ, and one has We note that the power µ, the effective equation of state parameter w, and the deceleration parameter q, all encode the same physical information.Similar to previously studied models, the positivity of the matter variable and equation (4.10) imply that 0 ≤ Σ ≤ 1, and hence 0 ≤ X 2 + Y 2 ≤ 1. Together with the fact that Y ≥ 0, since we are considering an expanding universe, this means that the phase space for the variables X and Y is a semicircle of radius one.
We are now ready to state the dynamical equations of the system, using the convenient variables defined in (4.11).This leads to two independent equations Here a prime denotes a derivative with respect to the logarithm of the scale factor log(a).One can now follow the standard dynamical systems approach to study this system, for a review see [50].We begin with the fixed points of (4.17)-(4.18).We note that these are two polynomial equations of degree three, meaning that one could find up to nine real distinct critical points, by Bézout's theorem.If Y = 0, the second equation is automatically satisfied, and this leads to the solutions X ∈ {−1, 0, +1}.Next, excluding Y = 0, one notes that Y appears only as Y 2 in the equations, meaning that there are up to four more solutions.Two of these are at negative values of Y , which we exclude, again because we are considering an expanding universe (Y ≥ 0).Assuming λ > 0 and −1 ≤ w ≤ 1, we obtain a total of five critical points, shown in Table 1.
Note that Point B is always located on the boundary of the phase space while Point C is generally inside the phase space, if it exists.For the special value the lower existence bound, Point C is also on the boundary.Next, one needs to study the eigenvalues of the stability matrix at each of the critical points.For more details about their classification, see Appendix A. For the first four points, O, A ± , and B, these are given in Table 2.Note that we will discuss the occurrence of possible zero eigenvalues separately to keep the discussion more straightforward.For example, one may immediately note that the choice w = 1 implies at least one zero eigenvalue for the Points O and A ± .The final critical point, C, is more difficult to study as the eigenvalues are much more involved.They are the solutions of the characteristic polynomial in ξ Solving this quadratic equation is easy, however, the explicit solutions do not offer much insight given that they contain three free parameters.For concrete parameter choices, we discuss this point in more detail below.One easy result to extract is the sum of the eigenvalues ξ 1 and ξ 2 at this point, that is, As this number is negative for w < 1, this point cannot have two positive eigenvalues and therefore will have at least one stable direction.This implies that Point C is a saddle point, stable node, or stable spiral.
There will be many parameter choices resulting in zero eigenvalues, w = 1 being the obvious one.However, the choice kλ = √ 6(1 + w) would also give a zero eigenvalue for Point O.The stability analysis of such points requires techniques beyond linear stability theory.These are well known and their applications in cosmology were discussed in [50,58,59].However, for the purposes of this work, we will assume a matter dominated universe w = 0 and employ linear stability theory.We note that our analysis can also be performed for the radiation dominated case w = 1/3 where one finds qualitatively similar results.
The matter dominated case
In what follows, we set w = 0. Points O, A ± and B are independent of w and all results discussed above apply.The location of Point C, if it exists, depends on w and so do its corresponding eigenvalues.We now outline some physical properties of the critical points of the system, with the values of effective equation of state parameter w and the deceleration parameter shown in Table 3.
In order to analyse the stability of the fixed points, we look at the different regions in the (k, λ)plane, see Fig. 1, and we recall that λ > 0. First, we remark that the fixed Points O and A ± exist for all values of λ and k.Moreover, there are four distinct regions of values of k and λ, which yield different stability properties of the critical points, and, hence, different cosmological phenomena.We discuss these four different cases and comment on their suitability as a cosmological model.2.
I II III IV
Region I.For values within Region I, there are only three critical points, namely O and A ± .In particular, O is a stable node, A − is an unstable node, and A + is a saddle.Since O is the only attractor of the system, all trajectories will eventually approach it.The Point A − can be thought of as the past-time attractor, in the sense that all trajectories would start at A − .Lastly, some trajectories are attracted towards A + , but are eventually repelled and move towards O.This case is not of physical interest.We do not show a phase space diagram.
Region II.In Region II, Point B does not exist.We therefore have four critical points: the unstable node at Point A − , the stable node at Point C, and the saddle Points O and A + .We note that here, Point C represents the scaling solution [60] as the effective equation of state parameter matches the matter one ( w = w = 0).Hence, the universe expands as if it was completely matter dominated despite the scalar field's influence, according to (4.13).We note that this is not an accelerated expansion but this solution is of physical relevance for the coincidence problem.The type of dynamics is illustrated by Fig. 2, where we set k = 1 and λ = 2.We remark the analogy with one of the cases discussed in [54], however we also point out that, in our example, no acceleration region is present.Region III.In Region III, there are five critical points in the phase space.Points A − , O, and C, still behave as an unstable node, a saddle point, and a stable node, respectively.Point A + is now an unstable node.Point B exists and is a saddle point.This is shown in Fig. 3, where k = 1/2 and λ = 3/2.Point C always lies outside the acceleration region, so it does not represent a late-time inflationary solution.This is, again, a scaling solution.We highlight the analogy with another case discussed in [54].All trajectories connect Points A ± to Point C, with the exception of the orbits along the boundary.
Region IV.In Region IV, there are again four fixed points since Point C lies outside the physical space.Here, Point A − is always an unstable node and can be seen as the past attractor.Similar to Region III, A + is an unstable node.Point O is a saddle point, whereas Point B is a stable node and therefore the late-time attractor.We note that here Point B lies within the region of accelerated expansion, hence we are in the presence of a cosmological solution with accelerated expansion.This is illustrated in Fig. 4. Once again, we emphasise the analogy with one of the cases discussed in [54].
A non-constant interaction model 5.1 Equations of the model
We are now considering a model with a non-constant interaction term.As we wish to exploit dynamical systems techniques without increasing the number of independent variables, we consider an interaction of the form [61,62] where α is a fixed power.Let us make the following observations to motivate this particular choice for f .In cosmology, s = s 0 and n = n 0 a −3 .Moreover, if we consider the linear equation of state 2) The dynamics of the system depends on the parameters w, λ, k, and α.We note that one could consider the limit α → 0 and recover equation (4.5).
It is clear that larger (integer) values of α can make the study of this system difficult, since the constraint (5.2) would become a polynomial equation of high order.At the same time, even the values α = ±1 introduce challenges as one has to deal with cubic equations.In fact, the two simplest cases that can be studied explicitly, without introducing further complications, are α = ±2.In what follows, we consider α = 2 and w = 0, and note that the radiation dominated case (w = 1/3) leads to broadly similar results.
When α = 2, the constraint (5.2) can be written as allowing us to eliminate σ from the equations, so that the dynamical system remains two-dimensional.Moreover, as σ 2 ≥ 0, we have This gives rise to the physical regions of the phase space 1 − x 2 − y 2 ≥ 0 and y 2 − kx > 0 , (5.5) 1 − x 2 − y 2 ≤ 0 and y 2 − kx < 0 . (5.6) Since the non-constant coupling leads to a considerably more complicated dynamical system, we restrict our study to the matter dominated case w = 0.The dynamical equation for x is given by where the functions A and B are defined by (5.9) Similarly, the y equation reads where (5.12) Here a prime denotes a derivative with respect to the logarithm of the scale factor log(a).This way of writing the dynamical equations, namely isolating the terms in powers of k, is useful as it allows us to consider the limit k → 0 easily.In that case these equations reduce to those of [54].
We remark that the acceleration equation, which follows from (3.2) and (3.4), is where the functions E and F are given by (5.15)
Critical points and stability
To find the critical points, we need to solve the equations x ′ = 0 and y ′ = 0 simultaneously, which is a non-trivial task since both numerators are polynomials of degrees seven, giving up to 49 roots.Many of those will lie outside the physical phase space, while others will come in complex conjugate pairs which also have no physical significance.At this point, it is not clear how many physical critical points this system will have for arbitrary λ and k and hence one has to investigate the system carefully to extract them.One way to find the critical points is to draw inspiration from the previous model.For example, setting y = 0 in (5.10) leads to y ′ = 0, while setting y = 0 in (5.7) means that x ′ = 0 simplifies to This yields the first set of critical points (−1, 0), (0, 0), (1, 0), and ( 3/2 /λ, 0).Secondly, we investigate critical points on the unit circle.By substituting x = cos θ and y = sin θ into (5.7) and (5.10), at a critical point we obtain This gives another critical point at (λ/ √ 6, 1 − λ 2 /6).Lastly, one can verify that setting x 0 = √ 3 ( √ 2λ) in the dynamical equations gives four additional solutions, other than y 0 = 0, which are We are not able to find other critical points in the physical phase space, either analytically or numerically.The critical points discussed above are summarised in Table 4, together with their corresponding value of the effective equation of state parameter and the value of the deceleration parameter.Note that there will be parameter regions where the critical points with y-coordinate y ± , called D ± , may not exist or where only one of these exits, see Fig. 5.
Table 4: Critical points of the dynamical system (5.7) and (5.10), for which an explicit expression could be found.We are now ready to investigate the stability of the critical points.This is straightforward for the Points O, A ± , B, and C. The result are collected in Table 5.
Point Eigenvalues Classification
Table 5: Stability properties of the critical points assuming λ > 0 for the non-constant interaction model.
For the Points D ± , the closed form expressions for the eigenvalues are very long and do not offer physical insight.However, when presenting specific cases, we give numerical values for the eigenvalues and discuss the various critical points in more detail.
Table 4 suggests that Point C is of particular interest to us.This point has an effective equation of state parameter w < −1/3 if λ < 15/2, and it is not located on the boundary of the phase space.For such a choice of λ, we note 15/2 > 3/ √ 2, which means this point will be an attractor of the dynamical system.In turn, such a model will naturally give rise to a period of late-time accelerated expansion.Moreover, since 15/2 > √ 6, Point B will not exists in this case.The physical phase space for these models is delineated by the upper semicircle of unit radius centred at the origin and the parabola x 2 = y/k, as described by (5.5) and (5.6).Subsequent figures will make clear which regions form the physical phase space.For all k, the semicircle and the parabola will intersect, creating two regions which meet at a point, and which trajectories can traverse.Note that the intersection point is, in general, not a critical point.
Phase space diagrams and physical interpretation
Different choices of λ and k result in rather different cosmological models, since the number of critical points and their location vary significantly.Below we consider several cases which illustrate the diversity of the dynamical behaviour exhibited by the model.We select values of λ and k systematically, but do not necessarily include every possible scenario which could arise in these models.
Case (i).We begin with λ = 3/2 and k = 8, as shown in Fig. 6.In this case, Point C does not exist, however, the other six critical points do exist.The phase space contains a region of accelerated expansion and we note that only Point O is in this region.Point O is an early-time attractor of the phase space, hence, this point could correspond to an early-time universe undergoing accelerated expansion.We note that this is negative but not less than −1/3, and therefore, not accelerating.Depending on the initial conditions chosen, some trajectories will approach Point D − with w = 0, which is matter dominated.On the other hand, trajectories starting at A + will either also terminate at B, or reach D + .This latter point is a stable spiral with w = 0, and hence corresponds to a matter dominated universe.It is interesting to note that trajectories in this case can briefly go through a region of accelerated expansion before reaching D + .While these parameter values yield an interesting phase space with a rich structure, this specific model has limited applicability for modern cosmology, since the stable fixed points do not lie within the accelerated region of the phase space.
Case (ii).Next, we consider the case λ = √ 3 and k = √ 8, see Fig. 7.In this particular case, Point D − coincides with Point B, this is true for all k.As in the previous case, Point C does not exist in the physical phase plane.Point O, as before, is an unstable node and acts as an early-time attractor.Point A − is a saddle and corresponds to the to other possible early-time attractor.According to Table 5, we note that Point B has eigenvalues 0 and −3/2, which means that we are dealing with a non-hyperbolic point.This point is a centre and one can verify that it is unstable.While this can be shown rigorously, it essentially follows from the fact that trajectories near B move towards the attractor D + , which is Case (iii).We will briefly comment on the case where λ = 3/ √ 2 and k = 2/ √ 3.For this particular choice, Points C and D − do not exist, while the Point D + is located at the intersection of the two regions of the phase space.This case is mathematically quite interesting, however, less so from a physical point of view.Of mathematical interest are the following facts: D + is a critical point of the system, however, both the numerators and the denominators of (5.7) and (5.10) vanish while giving a finite limit.The stability matrix is singular at this point, meaning that linear stability theory cannot be used.We will not discuss this case further.
Case (iv).Next, we consider the case λ = k = 1, which turns out to be physically interesting as a cosmological model, see Fig. 8.There are two unstable nodes, Points O and C, which act as earlytime attractors.As in the previous cases, Point O corresponds to an early-time universe undergoing accelerated expansion.Trajectories starting near O will eventually leave the acceleration region and be partially attracted to the saddle Point A − , after which they will reach the late-time attractor, the stable node at B. This point is also in the accelerated region, which means that this model not only allows for early-time acceleration (inflation) but also for late-time accelerated expansion.The effective equation of state parameter at B is w = −2/3, as can be seen from Table 4.All trajectories starting out in the right part of the phase space will also be attracted to B, making this the global attractor of the system.We remark that, in this sense, the dynamical behaviour is similar to that shown in Fig. 4.
Case (v)
. We now present the case λ = 1 and k = 20, see Fig. 9.Here all the fixed points we found analytically exist within the physical phase space.This model not only has a rich dynamical structure, but is also of physical relevance.We have two early-time attractors, Point O in the acceleration region, similar to the previous models, and Point C. Notice that Point C always satisfies w > 1, and so is not of physical interest.We are therefore most interested in trajectories starting from Point O.These will initially move towards A − , before leaving the left part of the phase space.By doing so, they will enter the acceleration region and move towards B where w = −2/3.Other than the various complications introduced by the other critical points, and the more complicated phase space structure, the physical situation is again somewhat similar to those shown in Fig. 4 and Fig. 8. Let us also mention that Point D + represents scaling solutions as the effective equation of state parameter is zero, and the universe evolves as if it were only matter dominated while also containing the scalar field.
Case (vi).To complete this section, we consider a case where the sign of the coupling is negative.We set k = −1/4 < 0 and λ = √ 5, and the phase plane is shown in Fig. 10.This model displays significantly different features than the cases where the coupling is positive.Point O can still be seen as an early-time attractor in the acceleration region.However, depending on the chosen initial condition, trajectories will either terminate at Point D + or Point C. Such trajectories can come close to the saddle Point D − , where w = 1/2.However, the effective equation of state parameter is also quite large at the other two points, meaning that one cannot have a model with a late-time behaviour close to a matter dominated universe.None of the late-time attractors appear in the acceleration region.Moreover, the left-hand side of the phase plane indicates the presence of a critical point at infinity,
Conclusions and discussions
The entire field of cosmology has seen remarkable progress in recent decades, including the use of dynamical system techniques to study the background behaviour of cosmological models.These techniques offer a systematic approach to understanding the underlying dynamics, which allows us to investigate the suitability of such models as realistic approximations of the universe.Our analysis involves mapping the cosmological equations onto a phase space, a step which relies heavily on the choice of suitable variables.This is rather non-trivial as various different variables could be employed and there is no particular reason to prefer one set of variables over any other.We therefore work with those variables that are known to be well suited for our task, see the review [50].
One of the main motivations of this work was to study models derived from a variational principle, in particular, we used Brown's approach for the formulation of the perfect fluid Langrangian for the cosmological matter.This approach allowed us to introduce new coupling terms, including boundary term couplings which have not been studied before in this context using fluids.So far, we have considered an algebraic vector coupling of the form f (n, s, φ)B µ J µ and noted that, in cosmology, one obtains the highly restrictive condition that f is proportional to n.Therefore, such a coupling does not yield interesting phenomena.We therefore focused on a derivative type coupling, f (n, s, φ)B µ ∂ µ φ, motivated by previous work [44].In that previous work, the coupling f (n, s, φ)J µ ∂ µ φ led to a minor change of the phase space compared to [54]: the critical point at the origin moves along the x-axis, depending on the choice of parameters.Our new coupling allows for a significantly different dynamical behaviour with features unseen before.Of particular interest to cosmology are situations where the model evolves through two periods of accelerated expansion, which replicate inflation and dark energy in a single model.
To gain an initial understanding of the resulting cosmological model, we began by studying the constant interaction model.This displayed similar behaviour to the well-known exponential potential quintessence model [54], where no interacting term is present.In fact, through a carefully chosen change of variables, we were able to arrive at a phase space which mirrors the one studied by those authors.These results demonstrate that our model can be seen as a natural extension of previous work.
We then proceeded to consider an interaction of the form n α/(2(1+w)) V −α/2 .This choice was motivated by the fact that couplings of this form will not increase the dimension of the phase space: it will remain two-dimensional.The key advantage of this assumption is that one can directly compare results with many previously studied models.In particular, we focused on the matter dominated case, w = 0, and chose α = 2.This choice leads to a rich dynamical structure with several distinct scenarios that can be of physical relevance.We were able to obtain solutions with early-time inflationary attractors, as well as late-time acceleration.Our models also included scaling solutions, which have received recent attention, see [63], as they may help to resolve the Hubble tension, that is, the discrepancy between the value of the Hubble constant inferred from measurements of the early universe and those derived from more recent observations [64,65].
Our approach to constructing coupled models lends itself to a significant amount of further study.First of all, one can study the constant coupling model in the radiation dominated universe.Our preliminary work suggests that the results are qualitatively similar to the matter dominated case, which is why we did not include them here, for the constant coupling model.One could attempt to present a comprehensive study for all w, however, this would not be without challenge as the convoluted equations would make analysing the stability at fixed points difficult.
Regarding the non-constant coupling model we proposed, there are three obvious extensions to our work, namely the cases α ∈ {−2, −1, 1}.For integer values of α, the phase will remain two-dimensional.However, one encounters other challenges which can be seen in equation (5.2).For example, when setting α = −2 one should eliminate y from the equation instead of the matter variable σ.This is unusual and has rarely been considered in the past.As a starting point, one would have to go back to the baseline model [54] and study it using a different choice of variables.In this way, comparisons could be drawn.For large values of α, on the other hand, one is dealing with a polynomial of a high degree, which is difficult to handle.In such cases, the best way forward would be to eliminate the variable x.This is equally unusual, and has also not been considered in the past.Given the complexities of these models, we are not able to predict the qualitative features of the resulting systems.The shape of the phase space alone changes significantly when varying the parameter α.
On top of all of this, one could, of course, drop the assumption of working with an exponential potential.It would be interesting to study our coupled models for power-law potentials or others.
Most of what has been done here will have to be re-investigated from scratch.For example, it is not even clear which types of couplings will admit a two-dimensional phase space.One would expect the constant coupling models to be similar to the uncoupled models, however, we refrain from speculating about results beyond this most basic of statements.
To determine the behaviour of trajectories near those fixed points, we can linearise the system around its critical point, by using a Taylor expansion for f in the neighbourhood of the fixed point.The dynamics of the linearised system are qualitatively equivalent to the original system.The eigenvalues of the matrix ∇f (x * ), known as the Jacobian matrix or stability matrix, contain the information about the local behaviour of f near x * .One generally speaks of stability or instability: • if all eigenvalues have positive real parts, we have an unstable fixed point or repeller • if all eigenvalues have negative real parts, we have a stable fixed point or attractor • if at least two eigenvalues have real parts with opposite signs, the corresponding fixed point is called a saddle point • if an eigenvalue is zero and at least one other eigenvalue has positive real parts, we have an unstable point • if an eigenvalue is zero and all other eigenvalues have negative real parts, linear stability theory does not suffice.
For more details, see for example [59].
Figure 1 :
Figure 1: Existence and stability regions in (k, λ)-plane.The plotted curves follow from the stability criteria shown in Table2.
Figure 2 :
Figure 2: Phase space with k = 1 and λ = 2. B is a stable node, that is, the only attractor describing a scaling solution with w = w = 0.No acceleration region present.
Figure 3 :Figure 4 :
Figure3: Phase space with k = 1/2 and λ = 3/2.Here the only attractor is Point C where the universe expands as if it is completely matter dominated (scaling solution), while Point B is a saddle point.The shaded region represents the area of the phase space where there is accelerated expansion.
Figure 5 :
Figure 5: Point D + exists in the whole shaded region.Point D − exists only in the dark grey region.
Figure 6 :
Figure 6: The parameter values are λ = 3/2 and k = 8.The eigenvalues of D + are −0.75 ± 1.3713i and the eigenvalues of D − are −2.1570 and 0.6570.The shaded area represents the part of the phase space where there is accelerated expansion.
Figure 7 :
Figure 7: The parameter values are λ = √ 3 and k = √ 8. B is an unstable centre and D + is a stable spiral.The shaded region represents the part where the phase space is accelerating.a stable spiral.It has eigenvalues −3/4 ± i √ 15/4.Similar to the previous case, Point O is an earlytime attractor corresponding to an early-time universe undergoing accelerated expansion.There is no late-time attractor within the acceleration region.
Figure 8 :
Figure 8: The parameter values are λ = 1 and k = 1.The shaded region represents the part where the phase space is accelerating.
Figure 9 :Figure 10 :
Figure 9: The parameter values are λ = 1 and k = 20.The eigenvalues of D + are −0.75 ± 1.11193i and the eigenvalues of D − are −2.34057 and 0.840574.The shaded area represents the part of the phase space where there is accelerated expansion.
Table 2 :
Stability of the critical Points O, A ± and B for system (4.17)-(4.18).The classification assumes that the eigenvalues are non-zero.
Table 3 :
Physical properties of the fixed points for the matter dominated case, for system (4.17)-(4.18). | 11,550 | 2024-04-08T00:00:00.000 | [
"Physics"
] |
DOMINO: a network‐based active module identification algorithm with reduced rate of false calls
Abstract Algorithms for active module identification (AMI) are central to analysis of omics data. Such algorithms receive a gene network and nodes' activity scores as input and report subnetworks that show significant over‐representation of accrued activity signal (“active modules”), thus representing biological processes that presumably play key roles in the analyzed conditions. Here, we systematically evaluated six popular AMI methods on gene expression and GWAS data. We observed that GO terms enriched in modules detected on the real data were often also enriched on modules found on randomly permuted data. This indicated that AMI methods frequently report modules that are not specific to the biological context measured by the analyzed omics dataset. To tackle this bias, we designed a permutation‐based method that empirically evaluates GO terms reported by AMI methods. We used the method to fashion five novel AMI performance criteria. Last, we developed DOMINO, a novel AMI algorithm, that outperformed the other six algorithms in extensive testing on GE and GWAS data. Software is available at https://github.com/Shamir‐Lab.
Introduction
The maturation of high-throughput technologies has led to an unprecedented abundance of omics studies. With the ever-increasing volume of publicly available genomic, transcriptomic, and proteomic data (Perez-Riverol et al, 2019), it remains a challenge to uncover biological and biomedical insights out of it. As data accumulated over the last two decades strongly indicate that the functional organization of the cell is fundamentally modular, a leading approach to this challenge relies on biological networks, simplified yet solid mathematical abstractions of complex intra-cellular systems. In these networks, each node represents a cellular subunit (e.g., a gene or its protein product) and each edge represents a relationship between two subunits (e.g., a physical interaction between two proteins) (reviewed in (McGillivray et al, 2018)). A biological module is described as a connected subnetwork of-molecules that take part in a common biological process. As such, modules are regarded as functional building blocks of the cell (Hartwell et al, 1999;Alon, 2003;Barab asi & Oltvai, 2004).
The challenge of identifying modules in biological networks, frequently referred to as network-based module identification or community detection, has yielded many computational methods (for a recent comparative study see (Choobdar et al, 2019)), and successfully identified molecular machineries that perform basic biological functions and underlie pathological phenotypes (Ideker & Sharan, 2008;Barab asi et al, 2011). However, such analysis is limited as it is based on a static snapshot of an abstract universal cell provided by the network, while the state of the cell greatly varies under different physiological conditions. One very powerful way to overcome this limitation is by integrating the analysis of omics data and biological networks. This approach overlays molecular profiles (e.g., transcriptomic, genomic, proteomic, or epigenomic profiles) on the network, by scoring nodes or weighting edges. This additional layer of condition-specific information is then used to detect modules that are relevant to the analyzed molecular profile (Mitra et al, 2013). A prominent class of such algorithms seek subnetworks that show a marked over-representation of accrued node scores (Ideker et al, 2002;Mitra et al, 2013;preprint: Reyna et al, 2020). Modules detected by such methods are often called "active modules," and following this terminology we refer to nodes' scores as "activity scores" and to the task of detecting active modules using such scores as Active Module Identification (AMI). (The task is sometimes called community detection with node attributes (Yang et al, 2014)). Hereafter, for brevity, where clear from the context, we refer to active modules reported by AMI methods simply as modules.
Modules detected by AMI algorithms are expected to capture context-specific molecular processes that correlate with the specific cellular state or phenotype that is probed by the analyzed omics profile (Mitra et al, 2013). Different AMI methods use different scoring metrics, objective functions, and constraints. For example, activity scores may be binary or continuous, the objective function could penalize for including low-scoring nodes, and constraints can limit the number of "non-active" nodes in a module. While the metrics by which modules are scored may differ from one method to another, the activity scores are always derived from the data (e.g., log 2 (fold-change of expression) for transcriptomic data). As the AMI problem has been proven to be NP-hard (Ideker et al, 2002), many heuristics were suggested for solving it (Mitra et al, 2013;Creixell et al, 2015).
Solutions reported by AMI methods comprise a set of active modules. A common downstream analysis is to ascribe each module some biological annotations that will point to the biological processes that it affects (Cerami et al, 2010;Leiserson et al, 2015;Barel & Herwig, 2018). This is most commonly done by testing enrichment of the modules for GO terms (The Gene Ontology Consortium, 2019). AMI solutions would ideally break down complex biological states into distinct functional modules, each mediating one or several highly related biological processes. For example, biological responses to genotoxic stress often comprise the concurrent activation and repression of multiple biological processes (e.g., DNA repair, cell-cycle arrest, apoptosis), each mediated by a single or a few dedicated signaling pathways (Ashcroft et al, 2000;Kyriakis & Avruch, 2012).
Another key advantage of AMI methods is the amplification of weak signals, where a reported active module comprises multiple nodes that individually have only marginal scores, but when considered in aggregate score significantly higher. This merit of AMI methods is especially critical for the functional interpretation of Genome-Wide Association Studies (GWASs) (Carter et al, 2013;Cowen et al, 2017). Numerous GWASs conducted over the last decade have demonstrated that the genetic component of complex diseases is highly polygenic (Khera et al, 2018;Musunuru & Kathiresan, 2019;Sullivan & Geschwind, 2019), affected by hundreds or thousands of genetic variants, the vast majority of which have only a very subtle effect. Therefore, most "risk SNPs" do not pass statistical significance when tested individually after correcting for multiple testing (Stringer et al, 2011;Boyle et al, 2017). This stresses the need for computational methods that consider multiple genetic elements together, to allow detection of biological pathways that carry high association signal. As a first step in this challenge, gene-level scores are inferred from the scores of the genetic variants that map to the same gene (de Leeuw et al, 2015;Lamparter et al, 2016). These gene scores then serve as activity scores by AMI methods for integrated analysis of GWAS data and biological networks. Recently, such analyses successfully elucidated novel process that are implicated in the pathogenesis of inflammatory bowel disease, Schizophrenia, and Type-2 diabetes (Chang et al, 2015;Nakka et al, 2016;Fern andez-Tajes et al, 2019).
In this study, we first aimed to systematically evaluate popular AMI algorithms across multiple gene expression (GE) and GWAS datasets based on enrichment of the called modules for GO terms. Remarkably, our analysis revealed that AMI algorithms often reported modules that showed enrichment for a high number of GO terms even when run on permuted datasets. Moreover, some of the GO terms that were often enriched on permuted datasets were also enriched on the original dataset, indicating that AMI solutions frequently include modules that are not specific to the biological context measured by the analyzed omics dataset. To tackle this bias, we designed a procedure for validating the functional analysis of AMI solutions by comparing them to null distributions obtained on permuted datasets. We used the empirically validated set of GO terms to define novel metrics for evaluation of AMI algorithm results. Finally, we developed DOMINO (Discovery of active Modules In Networks using Omics)-a novel AMI method and showed its advantage in comparison it to the previously developed algorithms.
Results
AMI algorithms suffer from a high rate of non-specific GO term enrichments We set out to evaluate the performance of leading AMI algorithms. Our analysis included six algorithms-jActiveModules (Ideker et al, 2002) in two strategies: greedy and simulated annealing (abbreviated jAM_greedy and jAM_SA, respectively), BioNet (Beisser et al, 2010), HotNet2 (Leiserson et al, 2015), NetBox (Cerami et al, 2010), and KeyPathwayMiner (Baumbach et al, 2012) (abbreviated KPM). These algorithms were chosen based on their popularity, computational methodology, and diversity of original application (e.g., gene expression data, somatic mutations) (Appendix Table S1). As we wished to test these algorithms extensively, we focused on those that had a working tool/codebase that can be executed in a standalone manner, have reasonable runtime, and could be applied to different omics data types. Details on the execution procedure of each algorithm are available in the Appendix. We applied these algorithms to two types of data: (1) a set of ten gene expression (GE) datasets of diverse biological physiologies (Appendix Table S2) where gene activity scores correspond to differential expression between test and control conditions, and (2) a set of ten GWAS datasets of diverse pathological conditions (Appendix Table S3) where gene activity scores correspond to genetic association with the trait (Methods). Note that for uniformity, we use the term activity also for the GWAS scores. In our analyses, we mainly used the Database of Interacting Proteins (DIP; (Xenarios et al, 2002)) as the underlying global network. Although the DIP network is relatively smallcomprising about 3000 nodes and 5000 edges, in a recent benchmark analysis (Huang et al, 2018), it got the best normalized score on recovering literature-curated disease gene sets, making it ideal for multiple systematic executions.
First, applying the algorithms to the GE and GWAS datasets we observed that their solutions showed high variability in the number and size of active modules they detected (Appendix Fig S1 and Appendix Fig S2). On the GE datasets, jAM_SA tended to report a small number of very large modules while HotNet2 usually reported a high number of small modules (Appendix Fig S1). jAM_SA showed the same tendency for reporting large modules also on the GWAS datasets (Appendix Fig S2). Next, to functionally characterize the solutions obtained by the algorithms, we tested the modules for enriched GO terms using the hypergeometric (HG) test with the genes in the entire network as the background set. Specifically, we used GO terms from the Biological Process (BP) ontology, using only terms with 5-500 genes. To avoid potential bias caused by the underlying network and datasets, we excluded from each GO class genes that were included in it based on physical interaction, expression pattern, genetic interaction, or mutant phenotype (GO evidence codes: IPI, IEP, IGI, IMP, HMP, HGI, and HEP). Next, as part of our evaluation analysis, we applied the algorithms also on random datasets generated by permuting the original gene activity scores uniformly at random. Notably, we observed that modules detected on the permuted datasets, too, were frequently enriched for GO terms (Fig 1A) Moreover, different algorithms showed varying degree of overlap between the enriched terms obtained on real and permuted datasets (Fig 1B). These findings imply that many terms reported by AMI algorithms do not stem from the specific biological condition that was assayed in each dataset, but rather from other non-specific factors that bias the solution, such as the structure of the network, the methodology of the algorithm, and the distribution of the activity scores.
A permutation-based method for filtering false GO terms
The high overlap between sets of enriched GO terms obtained on real and permuted datasets indicates that the results of most AMI algorithms tested are highly susceptible to false calls that might lead to functional misinterpretation of the analyzed omics data. We looked for a way to filter out such non-specific terms while preserving the ones that are biologically meaningful in the context of the analyzed dataset. For this purpose, we developed a procedure called the EMpirical Pipeline (EMP). It works as follows: Given an AMI algorithm and a dataset, EMP permutes genes' activity scores in the dataset and executes the algorithm. For each module reported by the algorithm, it performs GO enrichment analysis. The overall reported enrichment score for each GO term is its maximal score over all the solution's modules (Fig 2A). The process is repeated many times (typically, in our analysis, 5,000 times), generating a background distribution per GO term ( Fig 2B). Next, the algorithm and the enrichment analysis are run on the real (i.e., nonpermuted) dataset ( Fig 2C). Denoting the background CDF obtained for GO term t by F t , the empirical significance of t with enrichment score s is e(t) = 1-F t (s). EMP reports only terms t that passed the HG test (q-value ≤ 0.05 on the original data) and had empirical significance e(t) ≤ 0.05 ( Fig 2D). We call such terms empirically validated GO terms (EV terms). In addition, for each AMI algorithm solution, we define the Empirical-to-Hypergeometric Ratio (EHR) as the fraction of EV terms out of the GO terms that passed the HG test ( Fig 2E and F).
The DOMINO algorithm
Our results demonstrated that popular AMI algorithms often suffer from high rates of false GO terms. While the EMP method is a potent way for filtering out non-specific GO term calls from AMI solutions, this procedure is computationally demanding, as it requires several thousands of permutation runs. In our analyses, using a 44-cores server, EMP runs typically took several days to complete, depending on the algorithm and the dataset. Seeking a more frugal alternative that can be used on a desktop computer, we developed a novel AMI algorithm called DOMINO (Discovery of active Modules In Networks using Omics), with the goal of producing highly confident active modules characterized by high validation rate (that is, high EHR values).
DOMINO receives as input a set of genes flagged as the active genes in a dataset (e.g., the set of genes that in the analyzed transcriptomic dataset passed a test for differential expression) and a network of gene interactions, aiming to find disjoint connected subnetworks in which the active genes are over-represented. DOMINO has four main steps: 0 Partition the network into disjoint, highly connected subnetworks (slices). 1 Detect relevant slices where active genes are over-represented 2 For each relevant slice S a. Refine S to a sub-slice S' b. Repartition S' into putative modules 3 Report as final modules those that are over-represented by active genes.
Step 0-Partitioning the network into slices This time-consuming preprocessing step is done once per network (and reused for any analyzed dataset). In this step, the network is split into disjoint subnetworks called slices. Splitting is done using a variant of the Louvain modularity algorithm (Blondel et al, 2008) (Methods). Each connected component in the final network that has more than three nodes is defined as a slice ( Fig 3A).
Step 1-Detecting relevant slices Each slice that contains more active nodes than a certain threshold (see Methods) is tested for active nodes over-representation using the hypergeometric (HG) test, correcting the P-values for multiple testing using FDR (Benjamini & Hochberg, 1995). In this initial step, we use a lenient threshold of q-values < 0.3 to accept a slice as a relevant one ( Fig 3B).
Step 2a-Refining the relevant slices into sub-slices From each slice, the algorithm extracts a single connected component that captures most of the activity signal. The single component is obtained by solving the Prize Collecting Steiner Tree (PCST) problem (Johnson et al, 2000) (Methods). The resulting subgraph is called a sub-slice ( Fig 3C).
Step 2b-Partitioning sub-slices into putative active modules Each sub-slice that is not over-represented by active nodes and has more than 10 nodes is partitioned using the Newman-Girvan algorithm (Methods). The resulting parts, as well as all the subslices from step 2a of ≤ 10 nodes, are called putative active modules ( Fig 3D).
Step 3-Identifying the final set of active modules Each putative active module is tested for over-representation of active nodes using the HG test. In this step, we correct for multiple testing using the more stringent Bonferroni correction. Those with q-value < 0.05 are reported as the final active modules ( Fig 3E).
Systematic evaluation of AMI algorithms on gene expression and GWAS datasets
We next carried out a comparative evaluation of DOMINO, and the six AMI algorithms described above over the same ten GE and ten A Comparison of GO enrichment results obtained on the original CBX GE dataset and on one random permutation of the original gene activity scores of this dataset. The histograms show the distributions of GO enrichment scores obtained for the modules detected on the original and permuted datasets. The Venn diagrams show the overlap between the GO terms detected in the two solutions. B Comparison of GO terms reported on the original and permuted datasets. We used 1-Jaccard score to measure the dissimilarity between the GO terms detected on the two datasets. Values closer to 1 indicate low similarity (that is, lower bias). Each bar shows, per algorithm, this measure on the ten datasets, averaged over 100 random permutations. Datasets are ordered from left to right as in Appendix Tables S2 and S3. Dashed lines show the median score. A The AMI algorithm and the GO enrichment analysis are applied on multiple (typically, n = 5,000) permuted activity scores. B A null distribution of enrichment scores (-log10(pval)) is produced per GO term. C The AMI algorithm is applied to the original (un-permuted) activity scores, to calculate the real GO enrichment scores. D For each GO term, the real enrichment score is compared to its corresponding empirical null distribution to derive an empirical score. In this example, GO_3 passed the HG test, but failed the empirical test and thus was filtered out. E, F Distributions of HG enrichment scores for all the GO terms that passed the HG test and for the subset of the EV terms obtained on the SHEZH GE dataset by jActiveModules with greedy strategy (E) and NetBox (F). EHR measures the ratio between the number of EV terms and the number of GO terms that passed the HG test. The high EHR obtained by NetBox (close to 1.0) demonstrates the advantage of this algorithm in avoiding false terms. best with an average above 0.8. (Fig 4A and B). Importantly, these high EHR levels were not a result of reporting low number of terms: DOMINO reported on average more enriched GO terms than the other algorithms, except NetBox on GE datasets (Fig 4C and D).
Module-level EHR (mEHR)
While the EHR characterizes a solution as a whole by considering the union of GO terms enriched on any module, biological insights are often obtained by functionally characterizing each module individually. We therefore next evaluated the EHR of each module separately. Specifically, for each module, we calculated the fraction of its EV terms out of the HG terms detected on it (Methods). Modules with high mEHR score are the biologically most relevant ones, in the context of the analyzed omics dataset, while modules with low mEHR mostly capture non-specific signals. The comparison between mEHR scores obtained by the different AMI algorithms is summarized in Fig 5A. Notably, solutions can have a broad range of mEHR scores (for example, in NetBox solution on the IEM dataset, the best module has mEHR = 0.78 while the poorest has mEHR = 0). To summarize the results over multiple modules, we averaged the k top scoring modules (from k = 1 to 20; Fig 6A). In this criterion, DOMINO scored highest, followed by NetBox. The results for GWAS datasets are shown in Figs EV1 and EV2A.
A B Figure 5. AMI algorithms evaluated by the module-level EHR (mEHR) criterion on GE datasets.
A mEHR scores for each algorithm and dataset. Up to ten top modules are shown per dataset, ranked by their mEHR. Dot size represents module's size. The EHR column in green shows the number of EV terms and the number of significant terms found. B An example of a module from the solution reported by NetBox on the ROR dataset (mEHR = 0.88). The nodes' color indicates expression fold change (log scale) in the dataset. The black nodes are the network neighbors of the module's nodes. Nodes with purple border have significant activity scores (that is, significant differential expression; qval < 0.05). The EV terms for this module are shown in red and those that did not pass the empirical validation in blue. Furthermore, the EMP procedure enhances the functional interpretation of each module by distinguishing between its enriched GO terms that are specific to the real data (i.e., the EV terms) and those that are recurrently enriched also on permuted ones. This utility of EMP is demonstrated, as one example, on a module detected by NetBox on the ROR GE dataset (Fig 5B). This study examined roles of the ROR2 receptor in breast cancer progression, and the GO terms that passed EMP validation are highly relevant for this process (e.g., GO terms related to steroid hormone-mediated signaling pathways). In contrast, GO terms that failed passing this validation procedure represent less specific processes (e.g., "DNA-templated transcription, initiation").
Biological richness
This criterion aims to measure the diversity of biological processes captured by a solution. Our underlying assumption here is that biological systems are complex and their responses to triggers typically involve the concurrent modulation of multiple biological processes. For example, genotoxic stress concurrently activates DNA damage repair mechanisms and apoptotic pathways and suppresses cell-cycle progression. However, merely counting the number of EV terms of a solution would not faithfully reflect its biological richness because of the high redundancy between GO terms. This redundancy stems from overlaps between sets of genes assigned to different GO terms, mainly due to the hierarchical structure of the ontology. We therefore used REVIGO (Supek et al, 2011) to derive a non-redundant set of GO terms based on semantic similarity scores (Resnik, 1999;Lord et al, 2003). We defined the biological richness score of a solution as the number of its non-redundant EV terms (Methods). The results in Fig 6B show that on the GE datasets, DOMINO, and NetBox performed best. On the GWAS datasets, DOMINO performed best (Fig EV2B). Note that the interpretation of this criterion is condition dependent: High biological richness can be revealing or an indication of spurious results.
Intra-module homogeneity
While high biological diversity (richness) is desirable at the solution level, each individual module should ideally capture only a few related biological processes. Solutions in which the entire response is partitioned into separate modules where each represents a distinct biological process are easier to interpret biologically and are preferred over solutions with larger modules that represent several composite processes. To reflect this preference, we introduced the intra-module homogeneity score, which quantifies how functionally homogeneous the EV terms captured by each module are (Methods; Appendix Fig S3). For each solution, we take the average score of its modules. On the GE datasets, NetBox performed best (Fig 6C). On the GWAS datasets, DOMINO scored highest (Fig EV2C).
Robustness
This criterion measures how robust an algorithm's results are to subsampling of the data. It compares the EV terms obtained on the original dataset with those obtained on randomly subsampled datasets. Running 100 subsampling iterations and using the EV terms found on the original dataset as the gold-standard GO terms, we compute AUPR and average F1 scores for each solution (Methods). On the GE datasets, solutions produced by DOMINO and NetBox showed the highest robustness over a wide range of subsampling fractions (Fig 6D and E). On the GWAS datasets, DOMINO's solutions scored highest (Fig EV2D and E).
A breakdown of the evaluation criteria by their properties is shown in Fig 6F. Table 1 summarizes the benchmark results. DOMINO performed best on the GE datasets in five of the six criteria, and in all six criteria on the GWAS datasets. NetBox came second, performed best or timed for best in two criteria and second in the rest.
In addition, DOMINO ran much faster than the other algorithms, taking 1-3 orders of magnitude less time (Appendix Tables S4-S6). This speed allows to run DOMINO and the EMP procedure in reasonable time on a desktop machine. We also noticed that runtimes were markedly shorter on permuted datasets, probably since after permutation activity scores are spread more uniformly across the network, producing smaller modules.
Analysis of large-scale networks
Our benchmark used the highly informative but relatively small DIP network (~3k nodes and~5k edges) in order to allow systematic evaluation of multiple AMI methods on many datasets. Yet, much larger networks are currently available. To examine how DOMINO performs on larger network, we applied it on two state-of-the-art human networks: the HuRI network (8,272 nodes and 52,549 edges) (Luck et al, 2020) and STRING (with > 18K nodes and > 11M edges) (Szklarczyk et al, 2017). We also tested NetBox, the secondbest performer in our benchmark, on these larger networks. The edges of the STRING network are weighted with a confidence score, ranging from 0 to 1000, based on the strength of their supporting evidence. To make the execution of the EMP feasible, we kept only edges with score > 900. The resulting network had 11,972 nodes and 243,385 edges. Setting a running time limit of 5 hrs, DOMINO ◀ Figure 6. Evaluation results for the GE datasets.
A Module-level EHR scores. The plot shows the average mEHR score of the k top modules, as a function of k in each solution. Modules were ranked, for each solution, by their mEHR score. Then, for each solution with n modules we calculate the average mEHR of the top min(k, n) modules. Finally, we averaged the results and got the average mEHR of an algorithm. B Biological richness. The plot shows the median number of non-redundant terms (richness score) as a function of the Resnik similarity cutoff (Methods). C Intra-module homogeneity scores as a function of the similarity cutoff. D Robustness measured by the average AUPR over the datasets, shown as a function of the subsampling fraction. E Robustness measured by the average F1 over the datasets shown as a function of the subsample fraction. In (D, E), 100 samples were drawn and averaged for each dataset and subsampling fraction. F A breakdown of the evaluation criteria by their properties. Richness, EHR, and robustness score solutions based only on the whole set of the reported GO terms, without taking into account the results for individual modules. In contrast, mEHR and intra-module homogeneity score solutions in a module-aware fashion. From another perspective, biological richness and intra-module homogeneity consider the relations among the reported GO terms, while EHR, mEHR, and robustness do not.
completed all runs on both the HuRI and STRING networks, while NetBox did so on these two network for 8/10 and 2/10 of the GE datasets and for 9/10 and 9/10 of the GWAS datasets, respectively. Notably, DOMINO consistently outperformed NetBox on 23 of the 24 criteria on both networks and both types of datasets (Table 2). DOMINO also performed overall better when using different HG qvalue thresholds (Table EV1). Taken together, these results demonstrate that DOMINO maintains high performance when applied to large networks as well.
Analyzing the network contribution to non-specific GO enrichment bias Understanding the causes for over-reporting of enriched GO terms is a key question that arises from our study. One prominent potential cause is the network topology, as the modules sought are connected subnetworks, and connectivity also reflects functional similarity. To explore the contribution of the network to the GO enrichment bias, we next detected modules in the underlying DIP network without use of any condition-specific activity profile and identified the GO terms these modules were enriched for. Overall, 2,450 out of 6,573 (37%) BP GO terms were detected by this analysis, and we refer to them as net-terms. Notably, while net-terms were in general highly over-represented among the GO terms reported by AMI solutions (Fig 7A and B), these terms did not show higher rejection rate by the EMP procedure than the other BP GO terms (Fig 7C and D) (see Appendix for full details of this analysis). These results show that simple exclusion of GO net-terms from AMI analyses cannot replace the empirical validation to lessen over-reporting of non-specific GO terms. Better understanding of the bias origin is required.
Discussion
The fundamental task of active module identification (AMI) algorithms is to identify active modules in an underlying network based on context-specific gene activity profiles. The comparison of AMI A B C D Figure 7. Comparison of the GO terms identified by each benchmarked algorithm to the terms identified by using the network only (net-terms).
A, B Average number of net-terms and other terms. Only terms reported in four datasets or more were included. Note that no terms were reported in more than four GWAS datasets by DOMINO and NetBox, which obtained the best overall results ( Table 1). (A) GE; (B) GWAS. C, D Average rejection ratio of net-terms and other terms. The rejection ratio of a GO term in an algorithm is the fraction of datasets in which the term appeared as significant but was not empirically validated (see Appendix). (C) GE; (D) GWAS. P-values were calculated by comparing the rejection ratios between net-terms and other terms using Mann-Whitney U one-sided test.
algorithms is challenging due to the complex nature of the solutions they produce. Algorithms differ markedly in the number, size, and properties of the modules they detect. Although AMI algorithms have been extensively used for almost two decades (Ideker et al, 2002), there is no accepted community benchmark for this task and no consensus on evaluation criteria. As active modules are often used to characterize the biological processes that are most relevant in the context of the profiled activity, we analyzed the solutions produced by AMI algorithms from the perspective of enrichment for GO terms annotating biological processes. Previous works reported that the scheme used by the popular jActiveModule algorithm to score active modules is biased toward large modules and suggested ways to alleviate this bias (Nikolayeva et al, 2018) (preprint: Reyna et al, 2020. Our study reports on a different bias that is prevalent in AMI solutions: their tendency to report non-specific GO terms. Early on in our analysis, we observed that many enriched GO terms also appear on permuted datasets, suggesting that such enrichment stems from some proprieties of the network, algorithm, or the data that bias the results. To overcome this bias, we developed the EMP procedure, which empirically calibrates the enrichment scores and filters out non-specific terms. This procedure can be applied to any AMI algorithm. To exemplify their merits, studies that present a novel AMI method usually report a collection of enriched gene sets (e.g., GO terms or pathways) obtained on the algorithm's solution and are biologically relevant to the analyzed condition. While this approach is valid for demonstrating capabilities of an algorithm, it is problematic for a systematic evaluation of algorithms, due lack of gold-standard biasfree set of biologically relevant GO terms for a given condition. An additional difficulty is the hierarchical structure of GO ontology. A previous benchmark of AMI algorithms used as an evaluation criterion the fold enrichment of the output genes using a single set of biologically relevant genes (He et al, 2017). In our work, we defined five novel evaluation criteria based on the GO terms enriched in a solution, each emphasizing a different aspect of the solution (Fig 6F).
We used these criteria to benchmark six popular AMI algorithms and DOMINO, a novel algorithm we developed, on ten GE and ten GWAS datasets, which collectively cover a very wide spectrum of biological conditions. Overall, DOMINO performed best, indicating its ability to produce "clean," stable, and concise modules. NetBox also scored high in our evaluation. Interestingly, both DOMINO and NetBox use binary gene activity scores. One may expect that binarizing measured activity scores could degrade relevant biological signals. However, at least on our benchmark, binarizing the data helped in reducing noise and detecting modules that are specifically relevant for the analyzed conditions. Further study of this observation is needed.
Notably, the algorithms that we tested substantially differ in their empirical validation rates. Some algorithms produced solutions with very low EHR (< 0.5), and therefore running the EMP on them was critical. While empirical correction is desirable and adds confidence to the reported results, it is computationally highly demanding even with a relatively small network such as DIP. Naturally, using larger networks makes this procedure even slower (Appendix Tables S4--S6). A notable advantage of DOMINO is the high validation rates: On our benchmark, its average EHR and mEHR were above 0.84, suggesting that DOMINO can be confidently run without empirical validation when computational resources are limited.
A common caveat in any report comparing a novel method to extant ones is that the new method may be better tuned to the data than the other methods. This may introduce a bias in the reported results. In our case, we could not tune each of the other AMI methods due to the long running time of EMP. Community efforts like the DREAM challenges (Choobdar et al, 2019) help reduce potential bias by allowing authors to calibrate their own methods on a common set of test datasets. To enable additional testing, the code of DOMINO, EMP, and the evaluation criteria is freely available at https://github.com/Shamir-Lab/.
In summary, in this study we (i) reported on a highly prevalent bias in popular AMI algorithms, which leads to non-specific calls of enriched GO terms, (ii) developed a procedure to allow for the correction of this bias, (iii) introduced novel criteria for evaluation of AMI solutions, and (iv) developed DOMINO-a novel AMI algorithm with low rate of non-specific calls and better performance across most of the criteria.
The Louvain algorithm in DOMINO
The Louvain algorithm is a fast community detection method for large network (Blondel et al, 2008). This method aims to optimize an objective function by iteratively moving nodes between community to improve the objective function and fusing together the nodes of each community. In our benchmark, we used a variant (Lambiotte et al, 2008) that incorporates a resolution parameter denoted r, which we set to 0.15.
Threshold for testing relevant slices
Slices that contain only a few active nodes are unlikely to be relevant. Testing multiple such slices would diminish the significance of genuine relevant slices. Therefore, we test for relevance only slices that satisfy either.
#active nodes in slice #active nodes in network ≥ 0:1: or #active nodes in slice # nodes in slice ≥ α: where α ¼ min 0:7, #active nodes in network # nodes in network * 1 þ 100 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi # nodes in network p : The PCST application in DOMINO In PCST (Johnson et al, 2000), nodes have values called prizes, and edges have values called penalties. All values are non-negative. The goal is to find a subtree T that maximizes the sum of the prizes of nodes in T minus the sum penalties of the edges in it, i.e., ∑ v∈T pðvÞ À ∑ e ∈T cðeÞ where p(v) is the prize of node v, and c(e) is the cost of edge e. The node prizes are computed by diffusing the activity of the nodes using influence propagation with the linear threshold model ª 2021 The Authors Molecular Systems Biology 17: e9593 | 2021 (Kempe et al, 2015). The process is iterative: Initially, the set of active nodes is as defined by the input. In each iteration, an inactive node is activated if the sum of the influence of its active neighbors exceeds θ = 0.5. The influence of a node that has k neighbors on each neighbor is 1 k . Activated nodes remain so in all subsequent iterations. The process ends when no new node is activated. If v became active in iteration l then p(v) = β l where β ¼ max 0, 1 À 3 * #active nodes in network #nodes in network À Á . We define the penalty of edge e as c(e) = 0 if it is connected to an active node, and cðeÞ ¼ 1 À ɛ otherwise (we used ɛ ¼ 10 À4 ). PCST is NP-hard but good heuristics are available. In DOMINO, we used FAST-PCST (Hegde et al, 2014). The resulting subgraph obtained by solving PCST on each slice is called its sub-slice. See Fig 3C. The Newman-Girvan algorithm in DOMINO The Newman-Girvan (NG) algorithm is a community detection method (Girvan & Newman, 2002). This method iteratively removes edges using the Betweenness-centrality metric for edges and recomputes the modularity score for each intermediate graph. Let M i be the modularity score for the graph in iteration i. The process continues until a stopping criterion is met. The stopping criterion we used in step (2b) is log # of nodes in subÀslice ð Þ log # of nodes in network ð Þ ≤ M i .
Derivation of P-values and q-values for the GE and GWAS datasets
For the GE datasets, we calculated P-values for differential expression between test and control conditions using edgeR (Robinson et al, 2010) for RNA-seq and Student's t-test for microarray datasets.
We computed q-values using Benjamini-Hochberg FDR method (Benjamini & Hochberg, 1995). For GWAS, we used SNP-level Pvalues for association with the analyzed trait to derive gene-level association P-values using PASCAL (Lamparter et al, 2016), using the sum chi-square option and flanks of 50k bps around genes. We computed q-values using Benjamini-Hochberg FDR method (Benjamini & Hochberg, 1995).
Criteria for evaluating AMI solutions
We defined five novel criteria to allow systematic evaluation of solutions provided by AMI algorithms. For a specific solution, we considered the list of BP GO terms that passed the HG enrichment test (HG terms) and the terms that passed the EMP validation procedure (EV terms).
Solution-level criteria
Empirical-to-Hypergeometric Ratio (EHR) We define the Empirical-to-Hypergeometric Ratio (EHR) as the ratio between the number of EV terms and reported HG terms. EHR summarizes the tendency of an algorithm to report non-specific GO terms, with values close to 1.0 indicating good solutions while values close to 0 indicating poor ones. EHR reflects the precision (true-positive rate) of a solution.
Biological richness
This criterion quantifies the biological information collectively captured by the solution's EV terms. As there is high redundancy among GO terms-mainly due to the hierarchical structure of the GO ontology-we use the method implemented in REVIGO (Supek et al, 2011) to derive a non-redundant set of EV terms. This method is based on a similarity matrix of GO terms, which is generated using Resnik similarity score (Resnik, 1999). The biological richness score is defined as the number of non-redundant EV terms in a solution. We calculated this measure using different similarity cutoffs (1.0 to 4.0 in REVIGO).
Solution robustness
This criterion evaluates the robustness of a solution to incomplete gene activity data. It compares the EV terms obtained on the original dataset with those obtained on randomly subsampled datasets, where non-sampled gene levels are treated as missing. We repeated this procedure for subsampling fractions 0.6, 0.7, 0.8, and 0.9, iterating each fraction 100 times. Using the EV terms of the full dataset as the positive set, we computed average precision, recall and F1 scores across these iterations. Another perspective is provided by the examination of the frequency by which GO terms are detected in the subsampled datasets: higher frequency for a specific EV term implies higher robustness. We measured this robustness aspect of a solution using AUPR, in which EV terms are ranked according to their frequency across iterations (again, using EV terms detected on the full dataset as the positive instances). Note that cases in which an algorithm results in many empty solutions (that is, solutions with no enriched GO terms) and a few non-empty ones that are enriched for true EV terms can yield a high but misleading AUPR score. Therefore, we validated that the fraction of non-empty solutions obtained by the algorithms on the subsampled runs is high: All the algorithms achieved around 60% or more non-empty solutions on GE data (Appendix Fig S4).
Module-level criteria
Module-Level EHR (mEHR) This criterion calculates a single module's EHR. We define the module-level EHR (mEHR), as the ratio between the number of a module's EV terms and HG terms (Appendix Fig S3). We score each solution by averaging the mEHR of its k top-ranked modules (k values ranging from 1 to 20).
Intra-module homogeneity
This index measures the homogeneity of the biological signal that is captured by each module compared to the biological signal in the entire solution. For its calculation, we build a (complete) graph for the solution's EV terms (GO graph) in which nodes represent the EV terms and the weights on the edges are the pairwise Resnik similarity score (Appendix Fig S3B). Next, edges whose weight is below a cutoff are removed. The intra-module homogeneity is defined as the module's relative edge-density in this GO graph: of edges in module 0 s GO graph # of edges in a complete graph of that size # of edges in the solution 0 s GO graph # of edges in a complete graph of the same size We calculate the intra-module homogeneity score for a solution by averaging its modules' scores (Appendix Fig S3B). We repeat this test for a range of similarity cutoffs-from 1.0 to 4.0. This criterion provides a complementary view on top of the one captured by the biological richness criterion, by characterizing the biological coherence of the reported modules.
Data availability statement
The code for DOMINO, EMP, and the benchmark criteria is available at https://github.com/Shamir-Lab. The datasets used in this study are listed in the Appendix Tables S2 and S3. The GE datasets and the gene scores of both GE and GWAS datasets are also available at https://github.com/Shamir-Lab/EMP under the "datasets" folder.
Expanded View for this article is available online. | 9,703.4 | 2021-01-01T00:00:00.000 | [
"Computer Science"
] |
Lesser-Known Molecules in Ovarian Carcinogenesis
Currently, the deciphering of the signaling pathways brings about new advances in the understanding of the pathogenic mechanism of ovarian carcinogenesis, which is based on the interaction of several molecules with different biochemical structure that, consequently, intervene in cell metabolism, through their role as regulators in proliferation, differentiation, and cell death. Given that the ensemble of biomarkers in OC includes more than 50 molecules the interest of the researchers focuses on the possible validation of each one's potential as prognosis markers and/or therapeutic targets. Within this framework, this review presents three protein molecules: ALCAM, c-FLIP, and caveolin, motivated by the perspectives provided through the current limited knowledge on their role in ovarian carcinogenesis and on their potential as prognosis factors. Their structural stability, once altered, triggers the initiation of the sequences characteristic for ovarian carcinogenesis, through their role as modulators for several signaling pathways, contributing to the disruption of cellular junctions, disturbance of pro-/antiapoptotic equilibrium, and alteration of transmission of the signals specific for the molecular pathways. For each molecule, the text is built as follows: (i) general remarks, (ii) structural details, and (iii) particularities in expression, from different tumors to landmarks in ovarian carcinoma.
Introduction
There are several aspects which place the ovarian cancer in the focus of the scientific community. Its high mortality rate, due to the nonspecific symptoms that determine a delay of early diagnosis, the postsurgical treatment relapses, and the lack of favorable response to chemotherapy for most of the cases [1] require a better understanding of its mechanism and, implicitly, of the molecules that govern its behavior.
Although major progresses have been recorded in recent years in the knowledge of the complex signaling pathways involved in ovarian carcinogenesis [2], the deciphering of its pathogenic journey is far from being complete. The information on the genic and proteomic background of ovarian carcinoma (OC) could be regarded as a giant puzzle which is not yet assembled in order to form the entire image. On the basis of the molecular configuration of the signaling pathways, the interest of the researchers is focused on the identification of those components which could represent either new prognosis markers or new therapeutic targets, or both [3]. The difficulty of this endeavor is augmented by the histologic heterogeneity of ovarian tumors [4].
Even if in the last 15 years over 500 reports on the relationship between the molecular profile and tumor behavior [5,6] have been available in the mainstream publication, no new prognostic factor is yet confirmed and accepted. The ensemble of potential biomarkers in OC includes more than 50 molecules [5], from which the best known are WT1 and p53 (as oncogenes and tumor suppressor genes), Ki67, PCNA, and topoisomerase II (as proliferation markers), cyclins and their inhibitors (as cell cycle regulators), TRAIL and their receptors, Fas and Fas-L, Bcl-2, Bax, and caspases (as markers of apoptosis), BRCA and PARP-1 (as DNA repair enzymes), CD31, CD34, VEGF, COX-2, and MMPs (as angiogenesis markers), T lymphocytes and their regulatory protein (as immunological factors), EGFR and Her-2 (as tyrosine kinase 2 BioMed Research International receptors) and their signaling pathways, and cadherin-betacatenin complex [6]. Moreover, the review of the literature shows inconsistent data on other promising candidates.
Therefore, we believe the description of ALCAM, c-FLIP, and caveolin is worthwhile, because their expression is fewer investigated in OC, thus supporting their classification in the group of lesser-known molecules involved in ovarian carcinogenesis.
The choice of these three molecules with different functions is supported by our aim to illustrate diverse aspects of the events specific for carcinogenesis: disruption of cellular junctions, disturbance of pro-/antiapoptotic equilibrium, and alteration of transmission of the signals specific for the molecular pathways.
These molecules contribute to normal cell function, but their structural stability, once altered, reveals their competency as modulators that trigger the initiation of the carcinogenic mechanism.
The presentation respects the following sequences for each molecule: initial comments, structural features, and expression and known functions applicable in OC, with a corresponding discussion on the prognosis value.
Starting Point.
Cell-cell and cell-matrix interactions assist cellular differentiation and proliferation in both normal and pathologic development. Extensively investigated, the incomplete formation and/or remodeling of cell junctions are regarded as initial steps of the carcinogenic mechanism, while the detachment of cells from primary tumors sets in motion a course that favors invasion and metastasis. A particular attention is granted within this context to the cell adhesion molecules (CAMs), which comprise the families of integrins, cadherins, selectins, and immunoglobulin superfamily (IgSF). The organ specificity of the molecules belonging to IgSF (generically called Ig-CAMs) was studied in normal status and several malignancies [7][8][9][10][11][12][13][14][15][16][17][18]. For ovarian tumors, there is little specific information that ascertains the involvement of MCAM [19], L1CAM (CD171), EpCAM [20], IgLON [21], and ALCAM/CD166 (Activated Leukocyte Cell Adhesion Molecule) [22][23][24]. Strictly referring to ALCAM, besides its role of adhesion molecule, it is also a transductor that modulates a large panel of signaling pathways: MAPK, ERK1/2, and JNK [25].
Structural Features.
At first identified and isolated as ligand for CD6 [26] in thymic epithelial cells, ALCAM has been found since then in most fundamental tissues in the human body (except for muscle tissue) and in lymphohematopoietic structures. In physiological circumstances, ALCAM is involved not only in cell adhesion processes, but in neurogenesis, hematopoiesis, and immune responses as well [27]. The adhesion mechanism of ALCAM is both heterophilic (ligand-dependent) and homophilic (ligand-independent, regulated by actin cytoskeleton [28]) and is ensured either by interaction at the N-terminal domain or by cis oligomerization on cell surface through C-proximal domain [29].
Following the typical structural pattern of immunoglobulins, ALCAM is a type I transmembrane glycoprotein, with three domains: one extracellular (500 amino acids), one transmembranous (22 amino acids), and one short intracellular, cytoplasmic domain (34 amino acids) [30]. The extracellular domain consists of five N-terminal domains of immunoglobulin type; two are variable and three are constant (V1V2C1C2C3) [30]. The gene which codes ALCAM is located on the long arm of chromosome 3 [26].
ALCAM Expression: From Different Tumors to Landmarks in Ovarian
Carcinoma. In tumor pathology, ALCAM expression varies from strong (colon, gastric, and pancreatic cancer) [31][32][33] to weak (breast cancer) [34], depending on cellular type and on the modified microenvironment.
Unfortunately, as far as we know, although there are roughly 150 reports on ALCAM in various types of tumors, only one of these focuses on its value as prognostic factor in OC [23], based on the assessment of one human serous OC cell line and human tissue samples.
The role of ALCAM in ovarian carcinogenesis cannot be understood without knowing its behavior in the normal status. The multiple cell interactions promoted by ALCAM are due to the five extracellular binding domains Ig-like, which explain the membranous expression pattern revealed by immunohistochemistry (IHC). In malignancies, when intercellular adhesion is damaged, with loss of membranous contact, ALCAM expression relocates in cell cytoplasm. In other words, any loss of binding is associated with the internalization of ALCAM [23]. Hence, any event that perturbs the connection between ALCAM and its ligands brings about repercussions on the motility of ovarian tumor cells [23].
Thus, it is believed that the membranous expression of ALCAM reflects the maintenance of intercellular stability ( Figure 1) and that the cytoplasmic location, resulting from rearrangement of the intercellular junctions, characterizes tumor cells with high potential for invasion and metastasis [23]. This cytoplasmic specificity discriminates the advanced stages from the early ones, which designates ALCAM as a useful marker in the attempt to prove the effect of destruction of the intercellular binding, in tumor versus normal context [23]. Consequently, the decrease or absence of ALCAM membrane expression indicates a poor outcome in OC and can be useful in the identification of patients at risk, who need a more frequent follow-up and alternative treatment [23].
However, our experience in the IHC assessment of ALCAM expression in OC (unpublished data) revealed, in a completely unexpected manner, results that contradict the reports in the literature [23]. The membranous pattern of ALCAM, indicator of junction stability and, therefore, of low invasive potential, was predominantly associated with stage III and G3 differentiation. These results assign a higher potential for aggressiveness to the membranous pattern of ALCAM than the one generally recorded in tumor pathology and, particularly, in the ovarian malignancies. This statement opens a series of new perspectives for the reappraisal of the significance of ALCAM expression as indicator for tumor progression ability. In our opinion, a hypothesis worthy of consideration implies the return to the membranous expression, after the cytoplasmic translation, which would reflect a much more aggressive biological behavior than the cytoplasmic profile. Recent data relying on in vitro (using human epithelial OC cell lines) and in vivo (using human sera and ascites fluid) studies show the existence of a soluble form of ALCAM (sALCAM) [22,24], which results from its disconnection from the cell membranes ( Figure 1). EGFR, in association with other protein molecules (such as phorbol esters and pervanadate), via molecular signals triggered in various pathways, ensures the release of ALCAM from ovarian tumor cells through a metalloproteinases-dependent mechanism, regulated by the proteolytic activity of ADAM17/TACE, which determines the occurrence of sALCAM in ascites and serum [22,24]. Membranous detachment of ALCAM may also occur as result not only of protease degradation but also of methylation of ALCAM promoters [22,24]. sAL-CAM conducts tumor growth by coordination of invasion and metastasis [22,24]. The potential for diffusion in the extracellular liquid recommends the usage of sALCAM as ovarian tumoral biomarker [22,24], in correlation with the expression level, for sALCAM may be present in the serum of healthy individuals as well [42,45].
Structural
Features. c-FLIP functions as a complex multiprotein system consisting of 3 isoforms with roughly similar structures: a long variant c-FLIP L and two short ones, c-FLIP S and c-FLIP R [47,77,78]. The two short variants result from the nucleotide polymorphism in 3 splice site of c-FLIP gene [51,79] and are almost equal in size (26 and 24 kDa, resp.) and biochemical arrangement, with only one difference in the Cterminal domain, where c-FLIP S has an addition of 20 amino acids, essential for ubiquitination and proteasomal degradation, which support the antiapoptosis effects [46]. c-FLIP L is the longer variant, weighing 55 kDa, and has a structure similar to caspase-8, which it inhibits and deactivates. The structural analogy between c-FLIP and caspase-8 contributes to unfavorable effects with repercussions in cancer therapy [51]. All three c-FLIP variants display at their N-terminal end two death effector domains (DEDs) [51].
The c-FLIP protumoral effect is achieved by binding c-FLIP to the death receptors through DEDs (in a liganddependent or ligand-independent pattern), followed by inhibition of DISC formation by TRAIL and CD95/Fas/APO1 [49] and consequent blockage of the proapoptotic activity of caspase-8 and caspase-10, through inhibition of their activation [51] (Figure 1).
In ovarian carcinogenesis, the published data is centered on the antiapoptotic role of c-FLIP in the carcinogenic mechanism by using OC cell lines [49,71,72,76], while only four reports analyze its value as prognosis marker on human tissue samples [71,[73][74][75].
The presence of c-FLIP is associated with unfavorable prognosis [75], due to its contribution, by regulation of TRAIL signals [73], to the resistance towards the apoptotic receptors [74], which promotes ovarian tumor progression and development of chemoresistance [49] (Figure 1). However, although the knowledge on the various apoptotic receptors and pathways involved in sensitivity or resistance of OC to chemotherapy has increased significantly in the last two decades, this issue is still in permanent upgrade.
Our experience in the IHC appraisal of c-FLIP in OC (unpublished data) reveals, in accordance with the literature [71,[73][74][75], that the expression of c-FLIP varies significantly between the early and advanced stages, as well as in correlation with the differentiation degree. Our data indicates that a positive expression of c-FLIP characterizes the initial phases of ovarian carcinogenesis, which corresponds to FIGO I stage and differentiation degree G1.
The decrease of c-FLIP expression in advanced stages could be explained either by its interposition only in the initial phases of the apoptosis [74], this process being later inhibited by several other molecules which regulate tumor survival, or it could be possible that the intervention of c-FLIP is no longer necessary for the inhibition of the pathways involved in the maintenance of apoptosis.
It is worth mentioning that, in case of a functional p53, ovarian tumoral cells may escape from the cascade of events specific to apoptosis [71,75]. Inversely proportional relationships between c-FLIP and p53 are reported, with the c-FLIP increased expression being associated to "wild-type" p53, while mutant p53 is associated to diminished c-FLIP expression [71]. Consequently, the literature describes increased expression of c-FLIP in well-differentiated serous OC and clear cell OC, subtypes which, according to pathogenic classification, are type I tumors, without p53 expression at molecular level [71].
All these data recommend c-FLIP not only as a candidate prognostic factor for OC but also as an useful tool in patients' stratification for innovative treatments which could also take into consideration c-FLIP as therapeutic target [75,80,81].
They play a dynamic part in the mediation of intercellular and/or extracellular adhesion through cadherins, integrins [89], and fibronectin [90], in the control of endothelial passage, ensuring the stability of the endothelial barrier via catenins [91], and inhibit inflammatory processes, through their action on the cytokines [89].
Structural Features.
The caveolin is a transmembranous protein with heterooligomeric structure and a molecular weight of 24 kDa. The peculiar form of hairpin is caused by the organization pattern of its five domains: two cytoplasmic N/C-terminals, a C-terminal membrane attachment domain, an oligomerization domain, and a central transmembranous domain [87,89,112]. The oligomerized domain comprises BioMed Research International 5 a "scaffolding" subdomain ( Figure 1), responsible for the interaction between caveolins and various molecules in the vesicular traffic [87,87,112].
There are three types described: type 1 with two isoforms (1 and 1 ), type 2, and type 3, all with a molecular weight of 18 to 24 kDa [112]. For caveolins 1 and 3, the role as structural component of caveolae is ascertained, but the function of caveolin 2 remains still undetermined [113].
Caveolins are arranged in a regular pattern, with 100-200 molecules along a caveola, thus forming multiprotein complexes at the submembranous level [114]. Because of the numerous protein and nonprotein signaling molecules at these sites, any structural damage to the caveolae or caveolins generates the inhibition of molecular signaling [115].
Their involvement in the carcinogenic mechanism consists strictly in the regulation of signaling pathways Ras, Raf, ERK, ErbB-2-/MAPK/FAK, Src tyrosine kinase, PI3-K/AKT/mTOR, and NF-B [89,93,116], through their ability to block the activation of the oncogenes v-Src, H-ras, PKA, PKC, and Ras-p42/44 [94,113], and thus are granted the status of tumor suppressor genes [93,116] (Figure 1). However, recent evidence shows that caveolins can also act as oncogenes [117][118][119]. This potential duality, as oncogene versus tumor suppressor gene, reflects upon the different molecular pathways, which results in regulation of cell cycle, increase of tumor cell proliferation and invasion potential, promotion of angiogenesis, and the balancing of the apoptotic mechanism [93,117].
The little existing information regarding caveolins in OC is based rather on experimental researches [83,[106][107][108][109] than on human ovarian tissue specimens [110,111]. The first type of studies, on OC cell lines, shows that the caveolins have the same action mechanism as in the general sequence of carcinogenesis.
The IHC studies on paraffin-embedded samples of normal, benign, and malign ovary reveal that caveolins are present in normal ovarian surface epithelium, in benign pathology, and in early stages of tumor proliferation, with their expression being inhibited as the malignant transformation advances [111]. The prevalent association of the caveolins with the serous subtype is to be noted, in contrast with other OC histologic subtypes [111].
On the other hand, an increased expression of caveolins is ascertained in metastases, as opposed to primary ovarian tumors [110], a fact which suggests that caveolins should not be regarded merely as structural molecules, but also as functional ones, directly involved in the control and regulation of various signals that cross cellular membranes.
In accordance with the literature [110,111], our results in the assessment of caveolins in OC (unpublished data) indicate that absence of caveolin expression reflects tumor progression, and the correlations with clinicopathological factors and survival variables confirm that its negative expression is associated with a poor prognosis. Extrapolation of IHC results towards the mechanism that governs malignant transformation leads to the idea that in early tumor stages caveolins work as tumor suppressor genes, through the control of junctional contacts, while in advanced tumor stages caveolins function as oncogenes.
Hence, the role of caveolins in the mechanism of ovarian carcinogenesis remains to be clarified, more so taking into account the fact that their behavior varies, according to cellular microenvironment and received signals, from blocking the cellular oncogenic potential to stimulation of tumor growth [111].
Final Remarks
The current trend in ovarian carcinogenesis is the decoding of the genic and proteomic profile, which would lead to a deeper understanding of the pathogenic mechanism, a clearer explanation for the wide variability in the clinical course, and, also, to the documented validation of molecular markers with prognostic value.
This brief review of the three molecules, ALCAM, c-FLIP, and caveolin, chosen due to the interlocked dialogue they develop in the signaling pathways, is thus fully justified by the perspectives provided through the current limited knowledge on their role in the initiation and progression of ovarian carcinogenesis and on their potential as prognosis factors.
Conflict of Interests
The authors deny any conflict of interests. | 4,081 | 2015-08-03T00:00:00.000 | [
"Biology"
] |
Computational and Mathematical Methods to Estimate the Basic Reproduction Number and Final Size for Single-Stage and Multistage Progression Disease Models for Zika with Preventative Measures
We present new mathematical models that include the impact of using selected preventative measures such as insecticide treated nets (ITN) in controlling or ameliorating the spread of the Zika virus. For these models, we derive the basic reproduction number and sharp estimates for the final size relation. We first present a single-stage model which is later extended to a new multistage model for Zika that incorporates more realistic incubation stages for both the humans and vectors. For each of these models, we derive a basic reproduction number and a final size relation estimate. We observe that the basic reproduction number for the multistage model converges to expected values for a standard Zika epidemic model with fixed incubation periods in both hosts and vectors. Finally, we also perform several computational experiments to validate the theoretical results obtained in this work and study the influence of various parameters on the models.
Introduction
Every year over one billion people are infected from vectorborne diseases including malaria, Dengue, chikungunya, schistosomiasis, leishmaniasis, Chagas disease, Zika, and many more. These diseases affect urban as well as rural communities but thrive primarily among communities with poor living conditions. These vector-borne diseases also impose a substantial economic burden on families and governments. Hence understanding the spread of vector-borne diseases has been a major priority for many countries.
Over the years, many factors have contributed to the increase in the spread of vector-borne diseases including the ability of the vectors to adapt to new habitats, ability of the vectors to become drug-resistant, rapid human movement, and changes in policies on control measures. Mathematical models have often been employed to quantify such dynamics of the vector-borne diseases. These models are often described as compartmental model with the populations under study divided into compartments and, with appropriate assumptions, the different subpopulations transfer between these compartments. One of the earliest models was the formulation of a simple SIR model to describe the epidemic [1][2][3], where the entire population being studied was divided into a susceptible class ( ) which consisted of the number of individuals who are susceptible to the disease and are not infected at time , an infected class ( ) which consists of infected individuals who are assumed to be infectious and are able to spread the disease by direct contact with susceptibles, and ( ) which denotes the number of infected individuals who have been recovered and cannot spread the disease again. Most vector-borne diseases include an exposed phase ( ) between being infected and becoming infective. For a vector-borne disease this would mean a SEIR model for humans interacting with an SEI model for vectors as the vectors are not expected to recover in these models in the time span these models are solved. Currently, most models in the literature employ a SEIR-SEI single-stage model where the incubation periods are assumed to be fixed for both humans and vectors. Relaxing this assumption one may also employ a stage-progression model or the so-called linear chain trick that has been used for modeling diseases like HIV and Dengue [4][5][6][7] where the incubation may be modeled as the progression of multiple substages for humans and vectors. While these models have been considered for chikungunya and Dengue ( [8,9] and references therein), they have not been extended to Zika that also includes direct transmission until recently [10,11]. The single-stage model introduced in [10] was extended by [11] to include symptomatic and asymptomatic infectious stages for the human population along with effect of preventative measures.
Two mathematical quantities that are often of interest in these compartmental models is the basic reproduction number and the final size [12]. The basic reproduction number denoted by R 0 is defined as the number of secondary disease cases caused by introducing a single infective individual into a wholly susceptible population of both hosts (humans) and vectors (mosquitoes). Typically if R 0 > 1 an epidemic occurs while if R 0 < 1, there will likely be no outbreak. The value of R 0 helps to quantify the level of control intervention necessary to contain an outbreak. For example, in the case of malaria, a mathematical model was introduced in order to show that malaria can be greatly reduced by reducing the mosquito population density below a certain threshold [13]. There are multiple ways to mathematically estimate the reproduction number for vector-borne diseases [12]. However these estimates vary considerably that may be because of different external factors such as severity of disease, the level of public health surveillance, and local climate condition that can possibly affect the number of vectors and many other such external factors [14].
The final size of the epidemic refers to the number of members of the population who are infected over the course of the epidemic. While there are various approaches to obtain the basic reproduction number corresponding to the model being analyzed, there are no exact solutions for obtaining final size relations for vector-borne diseases. Recently, a final size relation for epidemic models of vectorborne diseases (that also included direct transmission) was obtained for an age of infection model that can be applied to Zika [14]. Specifically, this work derived an upper and lower bound for the final size relation. This model was formulated and analyzed considering infectivity depending on age of infection which allowed arbitrary periods of stay in each compartment and also the inclusion of control measures such as treatment, quarantine, or isolation. While this work provides a new insight to understanding the epidemics of vector-transmitted diseases through a final size estimate, the authors are not aware of any other work that establishes similar estimates with an upper bound and lower bound for a traditional SEIR-SEI vector-borne disease model that includes direct transmission.
For most vector-borne diseases such as Zika, there are currently no vaccines available and resistance to drugs is an increasing threat. Hence the CDC and WHO have recommended vector control as one of the essential ways to prevent disease outbreaks. One such intervention that has shown a lot of promise in vector-borne diseases such as malaria includes using insecticide treated bednets (ITN) which has been proved to be simple, efficient, and costeffective [15][16][17]. Using ITN can help reduce contacts between mosquitoes and humans at home by providing a physical barrier. The insecticide used to treat the bed net also repels mosquitoes ("excito-repellency"or "deterrence") thus increasing the personal protection offered by the net [18,19]. Finally, mosquitoes which are not repelled will most probably be killed as they come in contact with the insecticide as they often rest on the bed net after biting. For Zika there is a need to develop mathematical models that can help provide insight into the relation between increased coverage of ITN and the decrease of disease prevalence through a combination of the personal protection given by the repellency of the insecticide and the community protection given by its insecticidal action.
In this paper, we present the following new contributions for enhancing our understanding of the spread of Zika. First, we build on a single-stage model similar to what is considered in [10] and generalize them by including the impact of using selected preventative measures such as ITN in controlling or ameliorating the spread of the Zika virus. For this model, we derive the basic reproduction number and a sharp estimate for the final size relation. The derivation for the latter is a new alternative to the derivation of the age of infection epidemic model [14]. Specifically, we show that our result matches well with the results presented in [10,14] in the absence of any control measures. Next, we expand the single-stage model to a new multistage model for Zika that incorporates more realistic incubation stages for both the humans and vectors. For this model also we derive a basic reproduction number and a final size relation estimate for the first time. We observe that the basic reproduction number for the multistage model converges to expected values for a standard Zika epidemic model with fixed incubation periods in both hosts and vectors. This is because both the singlestage and the multistage models would be included in an age of infection model. The proof for the final size of the multistage model builds on the derivation for the single-stage model developed in the work and the result applies also to diseases that can be transmitted directly as well as through a vector. The work in this paper incorporates the multistage progression in the intrinsic incubation periods and can be extended to include extrinsic incubation periods as well. Finally, we also perform several computational experiments to validate the theoretical results obtained in this work and study the influence of various parameters in the model.
The outline of the paper is as follows. In Section 2, we present the mathematical framework used to study the transmission dynamics and control of the Zika virus during a single outbreak via a single-stage model. We derive the basic reproduction number and a new upper and lower bound estimate for the final size relation for the single-stage model that incorporates preventative measures such as insecticide treated bed nets. Section 3 carries out the basic analysis for an expanded multistage progression model that incorporates more incubation stages for the humans and the vectors. Finally, we present numerical results in Section 4 for both the Computational and Mathematical Methods in Medicine 3 single-stage and multistage progression models considered in the paper. Discussions and future work are presented in Section 5.
A Single-Stage Zika Model
In this work we develop and analyze an epidemic model for the spread of Zika through both a vector transmission and direct transmission via sexual contact. We will consider a constant total human population size ℎ with ℎ ( ) susceptibles, ℎ ( ) exposed, ℎ ( ) infected, and ℎ ( ) recovered.
Let the rate of Zika transmission through biting from mosquito to human be given in terms of the biting rate which corresponds to the number of bites in unit time. The effective mosquito bites in unit time that a susceptible human receives may be defined as the product of the biting rate and the probability ℎ that a bite transmits the infection which may also be referred to as infectiousness of mosquitoes to humans. Of this a fraction / is with an infective mosquito from ( ). Thus the number of new infective humans in unit time is ℎ ℎ ( / ). Defining the contact rate to be ℎ = ℎ / , the number of new infective humans in unit time is ℎ ℎ . The rate of the spread of Zika through direct sexual transmission from the infected human subpopulation to the susceptible human population is given by ℎ = ℎ / ℎ where ℎ is the sexual transmission rate of Zika. This adds to the number of new infective humans to be ℎ ℎ ℎ .
The vectors are assumed to move from the susceptible class ( ) to the exposed class ( ) through biting of an infected human. For the vectors, we consider a constant birth rate contribution of vectors in unit time and a proportional vector death rate in each of the susceptible ( ), exposed ( ), and infected ( ) vector classes, so that the total vector population size is constant. We will also assume that the vectors do not recover from infection and therefore there is no recovered class for the vectors. The total number of contacts by vectors sufficient to transmit infection therefore is ℎ and the corresponding vector transmission rate from the infected human to the vector is given by ℎ = ℎ / ℎ where ℎ is the infectiousness of humans to mosquitoes. Therefore, the number of new infective vectors in unit time is ℎ ℎ . We assume that the members of the exposed class ℎ move to become infectious at a human incubation rate of ] ℎ which is intrinsic human latent period. Members of the infectious human class recover with a rate of ℎ . Also, we let vectors of the exposed class move to become infectious with a vector incubation rate ] .
To incorporate preventative measures into the model, the effects of ITN are introduced in the rates of transmission from the susceptible human class to the exposed human class through a parameter measured as a percent = 1 − ITN. Note that when ITN = 1 (or = 0), the only movement from susceptible human class to the exposed class is through sexual transmission and not through the vector. On the other hand, if ITN = 0 (or = 1), the nets have no effect and the disease can spread through both vector and sexual transmission. As in the human model, we also incorporate preventative measure ITN into the vector model. We also introduce parameter for the removal of mosquitoes denoted by ℎ associated with ITN. To account for a wide range of behaviors, one can let the values of ITN from 0 to 1.
This leads to the following SEIR/SEI model for Zika transmission: where = ℎ ⋅ ITN. Note that the total human population ℎ appears implicitly in the parameters and there are no new births or deaths in the human population. Moreover, we will assume that the Zika epidemic is started by a visitor from outside the vector population . Hence we will assume (0) = and the total population of the mosquitoes and (0) = (0) = 0.
Derivation of the Basic Reproduction Number.
Recall that the basic reproduction number is defined as the number of secondary disease cases caused by introducing a single infective into a wholly susceptible population of both hosts (humans) and vectors (mosquitoes). Since the proposed mathematical model for human-vector interaction includes subpopulations with different susceptibility to infection, we will employ a general approach called the Next Generation Matrix approach [23][24][25] to find the basic reproduction number R 0 which is given by the following theorem.
Theorem 1. The basic reproduction number R 0 is given by
Proof. Given the infectious stages ℎ , , ℎ , in (1)-(7), we can create a vector F that represents the new infections flowing only into the exposed compartments. The components of the vector F are obtained by considering the terms denoting new infections from the susceptible equations (1) and (5) entering the exposed equations (2) and (6) with ℎ = ℎ and = N .
Along with F, we will also consider V which denote the outflow from the infectious compartments in (1)-(7) which is given by Next, we compute the Jacobian from F given by and the Jacobian from V given by Using matrices and one can then compute the Next Generation Matrix −1 given by ) .
Note that ( , ) entry of the Next Generation Matrix −1 is the expected number of secondary infections in compartment produced by individuals initially in compartment assuming that the environment seen by the individual remains homogeneous for the duration of its infection. Also, matrix −1 is nonnegative and therefore has a nonnegative eigenvalue. The basic reproduction number can then be computed as R 0 = ( −1 ) which is the spectral radius of the matrix. This eigenvalue is associated with a nonnegative eigenvector which represent the distribution of infected individuals that produces the greatest number R 0 of secondary infections per generation. In order to calculate the eigenvalues of −1 , we consider the characteristic equation where denotes the eigenvalues of the matrix and represents the Identity matrix. This can be simplified to yield The characteristic polynomial therefore is the following quadratic equation given by The basic reproduction number R 0 corresponds to the dominant eigenvalue given by the root of the quadratic equation ] . (17) Remark 2. Note that the infected human infects mosquitoes at a rate of ℎ ℎ / over an average time 1/ ℎ which produces ℎ ℎ / ℎ infected mosquitoes. Now a fraction ] /(] + + ) proceeds to become infectious. Next, the infected vectors infect humans at a rate of ℎ / ℎ for an average time of 1/( + ), producing ℎ / ℎ ( + ) infected humans per vector. The result is Also, sexual transmission produces ℎ cases in average time 1/ ℎ which then yields the additional reproductive number The basic reproduction number R 0 for system (1)- (7) can be written in terms of the basic reproduction numbers corresponding to the vector transmission R and direct transmission R as Note that Theorem 1 yields a general result for the basic reproduction number R 0 corresponding to the humanvector model given by equations (1)-(7) that include both sexual transmission and vector transmission. In the absence of one of these, the derived R 0 simplifies to physically meaningful mathematical quantities which are given in the next two corollaries.
Corollary 3.
In the absence of sexual transmission ( ℎ = 0), the basic reproduction number R 0 only corresponds to the vector transmission, that is, Corollary 4. In the absence of vector transmission ( = 0), the basic reproduction number R 0 only corresponds to the direct (sexual) transmission given by Note that in Corollary 3 (in the absence of sexual transmission), the next generation approach employed yields a square root in the reproduction number because it views the transition from humans to vector to humans as two generations.
Final Size for the Single-Stage Model.
In this section we will derive a relation between the basic reproduction number corresponding to the model equations (1)-(7) and the size of the epidemic. Note that the final size of the epidemic, the number of members of the population who are infected over the course of the epidemic, is − ℎ (∞) which is often described in terms of the attack rate (1 − ℎ (∞)/ ℎ ). We will first prove a lemma that will be used to derive the final size relation. (1)-(7), the total number of infected vectors depends on the dynamics of the epidemic and the total number of human infections as follows:
Next, we prove the following theorem that provides an upper bound for the final size for system (1)-(7) in terms of a basic reproduction numberR 0 that is closely related to R 0 that was derived in the previous section. (1)-(7) the final size relation can be bounded above as follows:
Theorem 6. For equations
where the basic reproduction numberR 0 is the sum of the sexual transmission reproduction number R and the vector transmission reproduction number R given byR 0 = R + R .
Remark 7. Note that integrating (1) and adding (1), (2), and (3) from 0 to , we get This leads to the form One can use this implicit relation between ℎ ( ), ℎ ( ), and ℎ ( ) to describe the orbit of solutions. In addition since the right hand side of (30) is finite, the left hand side is also finite and this shows ℎ (∞) > 0.
Next, we prove an estimate that provides a lower bound for the final size relation. The proof relies on the assumption that the vector population has a much faster time scale than the host population and therefore the vector population is at a quasi-steady-stage equilibrium, given by solutions to the equations for , , and in (1)-(7) that are constant functions of , but may depend on ℎ ( ), ℎ ( ), and ℎ ( ).
Theorem 8.
Let the vector population be at a quasi-steadystage equilibrium. The final size relation for (1)- (7) can then be bounded below as follows: where R * 0 is given by Proof. To determine the lower bound for log( ℎ (0)/ ℎ (∞)), we will first obtain a minimum for the Susceptible vector population . Since the vector population is assumed to be at a quasi-steady-stage equilibrium, we leṫ( ) = 0 in (5) which yields This can be rewritten to give since ℎ = ℎ / ℎ and ℎ ≤ ℎ . Therefore, Remark 9. Theorems 6 and 8 provide an upper and lower bound estimate for the final size relation, respectively, to yield Note that for = 1 and = 0, we are able to recover similar estimates that were derived for an age of infection epidemic model that included both vector transmission and direct (sexual) transmission [14].
To summarize, we have introduced the following variations of basic reproduction number in the description for the single-stage model: Remark 10. Note that in the derivation of the final size for the single-stage model we have assumed that (∞) − (0) = 0.
In general, however, one can expect (∞) − (0) = < 0 (as we expect the infections to die out). In this case, one can prove the upper bound estimate (30) as before; however, the lower bound estimate (41) is not a sharp estimate. This is not considered in this paper.
Computational and Mathematical Methods in Medicine 7
A Multistage Progression Zika Model
In this section we extend the single-stage Zika epidemic model to a multistage model by incorporating more realistic incubation period distributions. Specifically, we relax the assumption of fixed incubation period by using a stageprogression model or the so-called linear chain trick [4][5][6][7]. This Zika model incorporates incubation periods as the progression in ℎ incubation substages in humans ( ℎ 1 , ℎ 2 , . . . , ℎ ℎ ) and incubation substages in vectors ( 1 , 2 , . . . , ). This idea of a stage-progression model has been used for Dengue [7] which is caused by the same species of mosquito that causes Zika. The reason for introducing incubation substages is to model (for the first time for Zika) the time between a human is infected and the onset of symptoms due to the infection. These periods are important determinants of the temporal dynamics of the ZIKV transmission and are therefore critical for clinical diagnosis, outbreak investigation, implementation of prevention, programming control measures, and mathematical modeling. Under this formulation, the resulting incubation periods follow a gamma distribution with integer parameters ℎ and , respectively. When the rates of progression between substages are given by ℎ ℎ and for the incubation periods, the resulting gamma distribution has means 1/ ℎ and 1/ for the incubation periods, respectively, and the corresponding variances are given by 1/ ℎ 2 ℎ and 1/ 2 , respectively [7]. In this model we incorporate this stageprogression only in the intrinsic incubation period in humans and vectors. The analysis will be similar if one were to also incorporate stage-progression in the infectious period for the humans also.
We then have the following system of nonlinear differential equations describing the dynamics of Zika through a human-vector interaction aṡ In this section we prove a theorem that provides estimates for the final size for the system (48)-(56). First, we will first prove a lemma that will be used to derive the final size relation.
Theorem 12. For (48)-(52) the final size relation satisfies the following upper bound estimate:
where R 0 , the basic reproduction number corresponding to the system, is the sum of the reproduction numbers corresponding to direct (sexual) transmission R and vector transmissionR which are given by Proof. Adding (48)-(51) yields the following: This implies that ℎ ( )+∑ ℎ =1 ℎ ( )+ ℎ ( ) is a positive decreasing function and therefore the limit exists. The derivative of positive decreasing function tends to zero, and this yields that − ℎ ℎ → 0 and since ℎ > 0, this implies that ℎ → 0.
Integrating (74) we get Noting that ℎ (∞) = ℎ (∞) = 0 and ℎ (0)+∑ ℎ =1 ℎ (0)+ ℎ (0) = ℎ , (75) simplifies to the following: Employing (48): Substituting (76) and simplifying yields Integrating the left hand side and since the rate of sexual transmission ℎ = ℎ ℎ we get Substituting (57) from Lemma 11 into (79) gives us Computational and Mathematical Methods in Medicine 9 where we have used the integral mean value theorem as ℎ ≥ 0 on the interval [0, ∞) with min ≤ * ≤ max ≤ . Using (76) we can finally conclude that where we have used the fact that * ≤ . Define the reproduction numbers corresponding to direct (sexual) transmission R and vector transmissionR , respectively, to be we can now get the following upper bound estimate satisfied by the final size given by where R 0 = R +R . This proves the upper bound.
Corollary 13.
As the number of stages → ∞, the basic reproduction numberR corresponding to only vector transmission converges to the basic reproduction numberR * corresponding to a model that incorporates infected mosquitoes experiencing a fixed incubation period 1/] that is followed by an infectious state from which vectors do not recover. In particular, Proof. Using the definition ofR we havê where = ] /( + ).
Next, we prove an estimate that provides a lower bound for the final size relation. As in the single-stage model, the proof relies on the assumption that the vector population has a much faster time scale than the host population and therefore the vector population is at a quasi-steady-stage equilibrium, given by solutions to (48)-(56) that are constant functions of .
Theorem 14. Let the vector population be at a quasisteady-stage equilibrium. The final size relation for equations (48)-(56) can then be bounded below as follows:
where R * 0 is given by Proof. To prove the lower bound, recall that Also, since the vector population is assumed to be at a quasisteady-stage equilibrium, settinġ= 0 in (53) yields since ℎ = ℎ / ℎ and ℎ ≤ ℎ . Substituting (89) in (88) then gives (91) Remark 16. Note that in the derivation of the final size for the multistage model we have assumed that (∞) − (0) = 0.
In general, however, one can expect (∞) − (0) = < 0 (as we expect the infections to die out). In this case, one can prove the upper bound estimate (72) as before; however, the lower bound estimate (86) is not a sharp estimate. This is not considered in this paper.
To summarize, we have introduced the following variations of basic reproduction number in the description for the multistage model:
Computational Experiments
In this section, we validate our theoretical results developed in this work and perform simulations to predict the dynamics and estimate the basic reproduction numbers and final size relations. We implement the solution to the single-stage system (1)-(7) and the multistage system (48)-(56) in MAT-LAB using a fourth order Runge-Kutta method for solving ordinary differential equations. Specifically, the single-stage system of differential equations is solved using the script ode45 from MATLAB [26].
For our simulations, we considered the total human and vector populations to be ℎ = 1000 and = 4000, respectively, as in [14]. We considered one infective human initially and that the Zika epidemic is started by a visiting vector from outside the vector population . For the parameters in the models, we refer mainly to the data used in [10,14] that corresponded to the 2015 Zika outbreak in Barranquilla, Colombia. With the values for the parameters in Table 1, a linear relation was proposed in [14] given by We note that as the value of ℎ increases, the nature of the graphs also change. Specifically note that the final value of the number of susceptible humans ℎ ( ) decreases as ℎ increases.
Next, we consider the effect of the insecticide treated nets with no sexual transmission. So we let ℎ = 0 and let ITN = 0.1, 0.2, 0.3. The results are illustrated in Figures 4, 5, and 6, respectively. We clearly note that as the value of ITN increases in percentage (10%, 20%, 30%) of protection because of the nets, it takes more time for humans to get infected. Clearly this shows the effect of using nets.
Next, we consider the influence on the rate of sexual transmission ℎ on the various basic reproduction numbers that were obtained for the single-stage model. These include R 0 derived using the Next Generation Matrix approach in (20), the basic reproduction number corresponding to purely direct (sexual) transmission R defined in (19), the basic reproduction number corresponding to purely vector transmission R defined in (18), and the basic reproduction numberR 0 obtained in Theorem 6 which is the sum of R and R . Figure 7 compares these reproduction numbers as ℎ increases from 0 to 0.413. Next, we perform numerical experiments to validate the upper bound and lower bound estimate for the final size relation obtained in (46). This is illustrated in Figure 8 which clearly demonstrates the validity of the estimate. One may also note that log( ℎ (0)/ ℎ (∞)) is closer to the upper bound as pointed out in [14]. Also, for a purely direct (sexual) transmission,R 0 =R * 0 and therefore we get equality in the estimate that is denoted by the convergence at ℎ = 0.413 in Figure 8. Next, we considered the sharpness of the upper bound estimate by solving for ℎ (∞) in Specifically, we used Newton's method for solving nonlinear equations. The comparison of the results obtained from the simulations of the dynamics of (1)- (7) with the results obtained via Newton's method is illustrated in Figure 9. Note that the -axis values are normalized.
Since it is well known that increased coverage of ITN decreases vector prevalence, we performed our next computational experiment to explore the influence of using increasing values of ITN on the dynamics. The range of values included the absence of nets (ITN = 0) to completely protective nets (ITN = 1). The results are illustrated in Figure 10 which plots the final number of susceptible humans ℎ (∞), the basic reproduction numberR 0 , and the attack rate 1 − ℎ (0)/ ℎ (∞) for increasing values of ITN. The value of ℎ was chosen to be 0.2 for this simulation that corresponds to the inclusion of both vector and direct transmission in the model. As expected, Figure 10 illustrates the usefulness of insecticide treated bed nets to control the epidemic. The figure also can potentially help government officials to decide the level of control measures through insecticide treated bed nets that is needed in a certain area to contain the spread of the epidemic.
Next, the effect of employing a multistage model on the value of the basic reproduction number R 0 derived in Theorem 12 is illustrated in Figure 11 for increasing values of ℎ = 0, 0.1, 0.2, 0.3, 0.4, 0.413. As the number of stageprogression increases the value of R 0 converges for different values of ℎ . Finally, a convergence study was performed for the full multistage system (48)-(56) for increasing number of stages doubling each time, to demonstrate that the multistage model parallels the single-stage model. Figure 12 illustrates this convergence for fixed value of ℎ = 0.2. Note that the -axis values of the human population are normalized.
Discussion and Future Work
Over the last several decades, there has been an explosion in the development of mathematical models for outbreaks of infectious diseases that has become part of assessing epidemiological phenomena and making health policy decisions [1,12,27,28]. The outbreak of any new disease has always provided both an opportunity and a challenge for mathematicians and scientists. The opportunity leads to the development of improved models and the challenge is to make sure that these models represent reality. Zika is a great example of such as disease [29]. While there is a lot of published work and information available on models, methods, and simulations on infectious diseases that are purely vector-transmitted such as malaria, Dengue, and chikungunya and diseases that spread through direct transmission only such as influenza and AIDS, new diseases such as Zika that includes both vector transmission and direct transmission has provided a challenge for new mathematical models. For example, one of the essential mathematical tools in understanding disease dynamics is the calculation of the basic reproduction number that can help make informed decisions on whether there will be an epidemic or not. Another important quantity is the final size relation that provides a useful relation between the basic reproduction number and the size of the epidemic. While there has been a lot of progress made in computing these for various types of diseases, there is still a lack of complete understanding of these quantities for vector-borne diseases such as Zika that also includes direct (sexual) transmission.
Based on the approach we take, one can obtain different measures of the basic reproduction number for vector-borne diseases. However, there are no exact analytical solutions for the final size relation for such diseases. Nevertheless, one can obtain sharp estimates with an upper and lower bound for the epidemic size [14]. The latter formulated and analyzed a model with infectivity depending on age of infection. The work provided a useful upper bound and lower bound estimate for this age of infection model that applies to vector-borne diseases such as Zika that also includes direct transmission. In this work, we provide an alternate approach to determining similar estimates for the final size relation for enhanced SEIR-SEI single-stage and multistage progression models. The multistage model for Zika considered herein includes multiple incubation substates and was motivated by similar models for Dengue that did not account for the direct (sexual) transmission. Towards this end, we are able to successfully derive a new upper bound and lower bound estimate for the final size relation for the models considered. Moreover, we are able to show that the basic reproduction number for the multistage model proposed converges to the basic reproduction number corresponding to an equivalent nonlinear system of delay differential equations with fixed incubation periods in the humans and vectors.
Another contribution in this work is the inclusion of insecticide treated nets (ITN) that offer a mix of personal protection-blocking the bites of mosquitoes, thereby reducing the transmission from mosquitoes to humans-and community protection-reducing the longevity of mosquitoes and therefore the prevalence of the infectious stage of the disease, in mosquitoes. All the results developed in this work including the basic reproduction number and the final size relation estimate incorporate the influence of the ITN which are recommended by both the CDC and WHO as effective control measures.
We hope that the models, methods, and result from this work can help provide more insight into the propagation of a disease like Zika. The work also provides some opportunities for new avenues for future research. The multistage progression model in this work only considers multiple substates for the incubation for humans and vectors. However, one may also extend this work to also incorporate multiple infectious states in humans and compare such models against standard epidemic models with fixed incubation periods in both hosts and vectors and an exponentially distributed infectious period in hosts. One may also consider including indoor residual spraying (IRS) that can provide a coating of the walls and other parts of a house with a residual insecticide that can kill the vectors when they come in contact. One of the assumptions in this work involves employing a removal rate of vectors, and we need to assume vectors are in a quasi-steady state equilibrium depending on in order to continue to assume a balance relation for the host and vector contact rates. The parameter indicates the dependence of the model on bednets; it could also include a rate of killing of vectors by spraying, or we might wish to include separate control parameters for the effect of bed nets and the effect of spraying. Then an interesting question would be how the basic reproduction number and the final size of the epidemic depend on these control parameters. Finally, to model the effect of bednets, one may need some terms in the host equations that could resemble those in vaccination models. For example, this can include two susceptible compartments, with a rate of transfer from susceptibles without bednets to susceptibles with bednets who would have a smaller contact rate. In this case the rate of using bednets would be a control parameter. Finally, the derivations presented herein assumed that the initial population and the steady state population of the mosquito are the same. One can relax this assumption and as pointed earlier one can rederive the estimates. While the upper bound estimate can be derived in a similar fashion as in this paper, obtaining a sharp lower bound estimate will require some work. All these features and extensions will be considered in a forthcoming paper.
Conflicts of Interest
The authors declare that they have no conflicts of interest. | 8,628.2 | 2017-08-15T00:00:00.000 | [
"Mathematics"
] |
Evolution of the open-source data management system Ru-cio for LHC Run-3 and beyond ATLAS
,
Introduction
Managing large volumes of research data is a major challenge for any scientific project or experiment.The data requirements of these experiments are ever growing, leading to an unprecedented amount of required storage space and data organisation.These storage systems are typically heterogeneous and are distributed at multiple geographical locations under different administrative domains.The data workflows involved in managing and analyzing the data are also increasing in complexity, as the scientific data is usually produced, stored, and analyzed in multiple locations.
The ATLAS Experiment [1] at the Large Hadron Collider (LHC) [2] at CERN [3] is a typical example for such massive data requirements.To manage these large amounts of data, ATLAS developed the distributed data management system Rucio [4], which is a service taking care of the full data life cycle of the experiment.Rucio is now used in production at ATLAS, AMS [5], and Xenon1T [6] and is under evaluation by more scientific collaborations.
In this article we describe the evolution of Rucio to a generic, open-source, data management system for application beyond ATLAS.This article is structured as follows: In Section 2 we present the generic metadata support recently added to Rucio.We continue in Section 3 describing the plans for event level data management.Section 4 presents a preliminary study and plans about the increased usage of tape based storage systems and the support needed from the data management system.In Section 5 we present the plan for introducing new authentication and authorization methods beyond the current WLCG standard with X.509 certificates and proxies.Section 6 continues with new ways of deploying Rucio, specifically on Kubernetes, as well as the evolution of Rucio to a full-stack open source project.The article concludes in Section 7 with a summary.
Generic metadata support
Rucio already supports a limited set of metadata that can be associated with each data identifier.This includes system internal metadata (timestamps, states, access count, . . .), datamanagement metadata (checksums, file sizes, dataset length, . . . ) as well as a set of predefined user metadata (number of events, lumiblocknumber, task id, campaign, . . .).However, with the increased usage of Rucio in other experiments, even outside of the high-energy physics domain, it became very apparent that a more flexible way of managing metadata is needed.This is especially true for smaller experiments that do have the possibility to operate a dedicated metadata catalog next to their data management system.This development was guided by two design principles: The system must be able to add generic metadata to each data identifier.Thus, no pre-defined columns or data structures limiting the metadata are allowed.Secondly, registering a new data identifier including metadata should be an atomic operation.Thus it should not be possible that one part of the operation, i.e., registration of the data identifier or metadata, fails while the other is successful.Due to both requirements the design decision was made that the metadata will reside in the same relational and transactional database system as the logical data identifiers.This was mostly made possible due to recent advancements in Oracle [7], MySQL [8], and PostgreSQL [9] which allow the storage of arbitrary JSON datatypes and search operations within them.Thus the metadata associated to a data identifier is transformed into a JSON structure and stored to a JSON-type column within the relational database.
The development was done within Google Summer of Code 2018 [10] framework as part of the CERN & HSF project.APIs were created which allow a user to get, add/update, delete metadata for a specific data identifier.Also a command to search for data identifiers, matching a specific metadata filter within a scope was added.This new generic metadata feature is supported with Oracle 12c, PostgreSQL 9.3, and MySQL 5.7.8 and available with the Rucio 1.18.0 release.
Event level data management
The workflows of accessing data on the grid evolved during the last years.Today, particle physics production workflows access very specific physics events within files instead of entire files.This optimized and fine grained access of data will become even more prominent with LHC Run-3 and HL-LHC.This evolution is also expressed in recent developments with the ATLAS Event Service [11].
This trend raises the question when data objects within a file, such as events, become the common way to access scientific data, and if or how these objects should also have a dedicated representation within Rucio.There are essentially two options to realize this within Rucio.One is to expand the data identifiers (files, datasets, containers) to events and make sub-file-objects an official data identifier of Rucio.This is a major design change and would require changes throughout the full architecture stack.The advantage would be that users can operate on events similar to other data identifiers, thus add metadata to them, download them, or add replication rules.Next to the development effort the implications to the database layer would have to be evaluated, as storing sub-file information could result in one or two orders of magnitude more database objects.
The second option is to connect Rucio to an event catalog, such as the ATLAS Event Index [12].This would allow users to use Rucio tools to download specific events, as Rucio resolves the events via the APIs of the Event Index and then accesses the specific byte offsets of the file on storage.The development effort for Rucio is significantly smaller for this, but besides downloading specific events, the possibilities of interacting with events are only very limited.
It is not clear yet which option will be implemented, as it is mostly dependent on available development power and development plans of related projects.
Increased usage of tapes and networks
Rucio is prepared for an increased usage of tape in the experiment and is supporting the wider ATLAS Distributed Computing (ADC) tape carousel activity.The goal of the this first study within the tape carousel activity is to establish baseline measurements of current tape capacities, to know how much throughput we can expect from our current tape sites, stresstest to help optimize system settings, and identify bottlenecks.This study thus serves as the beginning of a decision process for the future developments and improvements required in Rucio with respect to hierarchical storage systems.
The actions of the iterative process are straightforward: run test, identify bottleneck, improve, repeat.The full test was run through Rucio, requesting the staging from tape to local disks at the sites.The data sample contained about 200 Terabytes of AOD datasets, with an average file size of 2GB.Replicas of the data that already existed on disk were manually cleaned beforehand.
The actual operation was then done in bulk mode, with three sites requesting a throttle of 2000 incoming staging requests.Two sites also had concurrent activities with production tape writing/reading and other experiment activity, but not at a significant level.
The tape frontend was quickly identified as the current bottleneck by limiting the number of incoming staging requests which were passed to the backend tape.This also limited the number of files which could be retrieved from tape disk buffer, and the number of files to transfer to the final destination.The requests which timed out were then rescheduled by Rucio, and eventually all requests were done.Figure 1 shows the summary of the individual tests.
Another problem that was identified was the size of the disk buffer on the tape pool servers, because dCache [13] reserves disk space before sending requests to backend tape and some sites may not have enough disk space in order to pass all requests to tape.The hardware solution is to increase this disk space to match the expected throughput from tape.Since we have not yet identified the full workflow for the tape carousel activity, this is unrealistic.The software solution is to loosen the reservation requirement or make it configurable and will be discussed with dCache developers for a stop-gap solution.Since we cannot expect sites to increase hardware an alternative can be applied from the Rucio side.Rucio has the capability to throttle transfer requests, and this could be extended to tape systems with knowledge about the buffer size.That way we can ensure that custom throttling can be applied per endpoint.Some pool servers also got overloaded with thousands of requests, it is unclear yet why, but since dCache keeps track of status of all the staging requests submitted to the backend tape, the retrievals from tape, once they are ready, and eventually transfer them to the final destination, a capacity upgrade of the number of pool servers seems almost mandatory.A potential software improvement could be in the mechanism of tracing requests and retrieving them when ready only.
Another finding was the distribution of requests among pool servers, where requests got dispatched to dCache pool servers that have more space.Eventually requests got stuck on low performing servers, although they have more space.Manual intervention was needed to redistribute the requests to other servers.There's a clear software improvement possible in the way requests can be dispatched among pool servers and automatically recovered when stuck.
For writing to tape, good throughput was seen from sites who repack or organise before writing to tape.This was the major reason for performance difference between sites with similar system settings.Since this requires writing in the way one wants to read later, it is not straightforward, especially if there are multiple workflows to support.Even though tape systems provide such file families and most sites use it, a grouping by experiment dataset seems more convenient and efficient.For full tape reading, near zero remounts were observed with sites that are already doing that.Finally, ADC is also working on bigger file size with a target of 10 Gigabytes as this will greatly increase tape writing throughput.
Next to tapes, another avenue for R&D is the network being used by Rucio.Especially the orchestration of transfers across wide area networks has received little attention yet due to the overabundance of network capacity.We foresee a major increase of network use therefore several activities have started to evaluate the use of novel network features.At the time of writing we have recently finished the first preliminary test of bandwidth load balancing.
The goal of the test was to overload the LHCOPN [14] links between CERN and SARA-NIKHEF with fake Rucio file transfers, and then use manual router loadbalancing to offload traffic from LHCOPN to LHCONE to increase the overall available bandwidth.We want to verify that the offload is effective and thus it is possible to use spare bandwidth when the primary links are congested.Eventually, the goal is that this router loadbalancing will be triggered by Rucio automatically.
We identified a data sample of 30 Terabytes, containing 4500 files, and estimated roughly 4 hours transfer time.Shortly after the transfer start the two links were fully used at 10Gbps each, however after half an hour the throughput went down to 60 percent without having activated the additional bandwidth.It was quickly discovered that the source storage system was badly configured in their loadbalancing algorithm which caused hotspots on the disk pools.After the intervention, the throughput increased again to 80 percent usage.
Unfortunately it was not possible to saturate the links, therefore the activation of the additional bandwidth did not show any increase of overall transport throughput, since the primary link could not be overloaded.The destination storage did not exhibit any abnormal characteristics in the meantime.Another unexpected fact was that most of the traffic was on IPv6 and not IPv4, which required some fixing of the routing tables configuration on the CERN outgoing routers.
Eventually the test was aborted and will be repeated in the future with greatly scaled out source storage to ensure link overload.Having this alternate path mechanism automatically triggered through scheduling decisions in Rucio promises to be extremely beneficial for large flows which are typical for experiment data export.
New authentication mechanisms
Across the WLCG the primary method of authentication and authorisation are X.509 certificates and their derivative proxies.While X.509 certificates are an industry standard they come with administrative and technological processes, such as certification authorities, which can be inhibiting to new experiments, smaller data centres, as well as sometimes the software itself.Keeping certificate authentication and authorisation synchronised across an experiment is also a source of major configuration and operational troubles.Within Rucio, the X.509 layer is one of the primary supported method next to username/passwords, Kerberos, and SSH public key exchange.To support a wider community of experiments new ways of authentication are in development, most importantly capability-based tokens and OAuth2/OpenID workflows.Both are expected to go into user testing during 2019.
Macaroons [21] are similar to HTTP cookies with several added benefits: while they can be delegated like cookies, they allow cryptographically-secure attenuation to restrict capabilities, and most importantly can be verified by a global generic verifier.Macaroons can thus be seen as the authorisation equivalent of symmetric cryptography.Alternatively, SciTokens [15] take the opposite approach, where instead of restricting from a full credential, claims are added to the bearer token giving additional rights.Verification follows the asymmetric cryptography approach, where the verification keys are publicly available.
The OpenID authorisation workflow built on top of OAuth2 is de-facto required for seamless integration with other federated infrastructures such as EduGAIN, and also serves as the distribution mechanism for the Macaroons and SciTokens.Rucio would create the necessary bearer token of the appropriate type depending on the request of the client.The tokens can then be handed off to FTS [22], which would use the restrictions or claims to authorise access to storage.Along the same lines, the tokens can be distributed to users who can then interact directly with the storage.Rucio could thus take the rule of what is the certification authority in X.509 environments, saving a full layer of authorisation and authentication and reducing individual node management complexity.
Deployment and open source software development
Rucio is available on PyPi [16] in three different packages: clients, webUI and a general package including server and daemons.On Docker Hub [18] we provide different containers for server, daemons, clients, and webUI which are heavily used in different deployments.The supported Python versions for the clients are 2.6, 2.7 and 3, while the server package is Python 2.7 compatible.Full-Stack Python 3 compatibility is planned for early 2019.
Current deployment of Rucio for ATLAS relies on Puppet to deploy Rucio Python packages and configurations on a multitude of different virtual machines.A large number of machines is used to provide redundancy and to separate different workflows, e.g., use different servers for the WMS and users.This has proven to work well in the ATLAS use case but it also results in a low utilization of the VMs and a complicated deployment for new experiments.For this reason a Kubernetes [17] deployment is currently evaluated.Kubernetes is a orchestration system for automated deployment, scaling and management for containerized application.For Rucio two separate containers for the server and daemons components are available.The containers are automatically build with every new release and openly available on Docker Hub [18].They are providing an easy way to run Rucio in a controlled environment without the need to manually install Python packages with possible dependency issues.On top of that Kubernetes simplifies the deployment and scaling of those containers.Kubernetes usually runs on a set of nodes where one is the master and the rest are minions.The minions run the actual containers and the master is used to evenly distribute the work on the minions.The operator has to define a set of different Kubernetes resources to run the Rucio servers and daemons and to also make the servers available for outside access.Large parts of the resource setup can be reused and here is where Helm [19] comes in.It is a package manager for Kubernetes similar to apt on Debian or yum on Red Hat.Packages for the servers and the daemons are available on Github [20].These charts (packages) can be installed with one simple command and the operator only has to provide a configuration file including information like the DB connection string, server/daemons configurations and the number of containers to run per service.
As mentioned before, this type of deployment is currently evaluated.There is a test cluster for ATLAS available that currently runs two integration servers and one integration daemon (Rule evaluator).Furthermore, a complete new instance of Rucio has been setup only using Kubernetes/Helm for the WLCG Data Organization Management Access (DOMA) third party copy (TPC) working group to initiate a constant flow of TPC transfers between test endpoints.This cluster runs in a stable manner providing valuable operational experience that will be useful later for production instances.
The Rucio project is organized as an open-source development project, under Apache 2.0 license, with contributions coming from 30 different contributors, with 19 contributors from the ATLAS collaboration.The establishment of the project as a community project, within the high-energy physics domain and beyond, was culminated with the hosting of the first Rucio Community Workshop in March 2018.Over 90 participants coming from 16 different communities participated in the workshop and presented their use cases and data management needs.The development is conducted as a formal software engineering process with the objective to produce performant, well-documented, scalable, and sustainable code.The Rucio development team conducts weekly development meetings with issues and development plans well documented on Github [23].There is human code review of all pull requests and automatic unit testing with more than 400 unit tests with different Python version against different database backends.We plan to foster this community approach by organizing yearly community workshops and fully focusing on the open-source development process of the software.
Summary & Conclusion
Managing large volumes of scientific data is a major challenge for experiments.Rucio provides a well established solution for these problems and has a demonstrated track-record of stability, scalability, usability, and performance.In this article we present the evolution of Rucio as an open-source data management system for LHC Run-3 and beyond ATLAS.The current development plans focus on the features needed for LHC Run-3.However, due to the recent interest of research experiments within the HEP community and beyond, plans also include features relevant for these new communities.Generic metadata support is one of these features mostly relevant for new communities.Previously, Rucio only supported a limited amount of metadata, but with this new development, based on JSON columns in state of the art relational databases, Rucio fully supports generic metadata within the transactional system.Event level data management is a planned development for LHC Run-3 and HL-LHC.The design plan foresees two options of including event information into Rucio and therefore making events directly addressable within Rucio.A first study for the increased usage of tapes and networks in the tape carousel activity have shown great promise of more widespread use of tapes.Rucio is fortunately already prepared and can serve as the source for efficient tape orchestration through site distribution and dataset organisation.The first network routing tests also showed promise even though we were not able to saturate the links due to source storage overload.These tests will be repeated in the near future.The current industry standard to authenticate and authorize users to the data management system and to storage are X.
509 certificates.To support a wider community of experiments, new ways of authentication are in development, most importantly capability-based tokens and OAuth2/OpenID workflows.Both are expected to go into user testing during 2019.Rucio is available on the Python Package Index and as containers on Docker Hub.Recently Kubernetes and Helm deployments are being evaluated with the objective of supplying a turn-key deployment for new experiments via Helm.Rucio is organised as an open-source project, under Apache 2.0 license, and recently fully embarked on the community path by hosting the first Rucio Community Workshop.The development team is now adapting to this community efforts to open up the development process and integrate use cases and developments beneficial to other experiments.
Table 1 .
Chart displaying the different performance characteristics of the measured tape systems. | 4,583.6 | 2019-07-01T00:00:00.000 | [
"Physics",
"Computer Science"
] |
Designing of Web-Based Learning Media for Senior High School During The Covid-19 Pandemic
. The learning process of Senior High School level is done by various methods. In the current pandemic condition, a learning process is needed that is implemented with strict health protocols. One alternative method of the effective and online learning processes is by learning to use web-based learning media. The creation of learning media in this study uses a prototype model, which consists of listening to needs, building mock-ups and evaluating mock-ups. As a sample of research is SMAS Wiyata Mandala Balai Batang Tarang, with data collection techniques used consisting of interviews, observations and library studies. The media designed in this study provides facilities to three (3) level users, namely administration, teachers and students. Administration can manage administrative data, majors, subjects, classes, students, teachers, teaching teachers, announcements and access student achievement reports. Teachers can manage to manage teaching materials, problem banks, meeting schedules and access student achievement reports. Students can access modules, work on weekly meeting questions and access the grades in the meeting results. This designed learning media is expected to help SMAS Wiyata Mandala Balai Batang Tarang in carrying out the learning process and can improve student achievement even in the midst of the pandemic.
INTRODUCTION
The condition of the 2019 corona virus disease (Covid-19) pandemic which has hit almost all countries throughout the world, presents its own challenges to the education sector, especially Primary and Secondary Education. In fighting the Covid-19 outbreak, the Government has implemented Health Protocols including prohibiting crowding, social distance, maintaining physical distance (physical distancing), wearing masks and getting used to always wash hands. 19), in general changes: Face-to-face learning in educational units in the 2020/2021 academic year and 2020/2021 academic year is carried out in stages throughout Indonesia with the following conditions: a. Education units located in the GREEN and YELLOW ZONE areas based on data onto the National COVID-19 Handling Task Force (https://covid19.go.id/petarisiko) can conduct face-to-face learning in education units after obtaining permission from the local government through the local government agency. provincial or district / city education, regional offices of the provincial Ministry of Religion, and district / city Ministry of Religion offices according to their respective authorities based on the approval of the local task force to accelerate handling of COVID-19; b. Education units located in the ORANGE and RED ZONE areas based on data onto the National Covid-19 Handling Task Force, are prohibited from carrying out faceto-face learning processes of educational units and continue to learn from home (BDR) activities. [1] Learning is a teaching and learning activity, which is usually only done face-toface. With advances in information technology, learning activities can also be carried out through the media, namely the website, which will make the learning process more efficient, effective, and flexible so that it is not fixed on face-to-face.
Learning is a learning and teaching activity that occurs directly, where there is an interaction between students and teachers. Learning activities carried out by students can encourage change in students, both to improve student knowledge, understanding, and skills which are constant. [2].
Learning is an activity carried out between two actors, namely teachers as educators and students as students, educators, namely actors who provide teaching material while student actors are recipients of teaching materials. [3]. It is concluded that learning is a learning and teaching activity that is carried out consciously between students and teachers, where the parties interact which shows the process of delivering and receiving teaching materials with predetermined targets. This learning is able to encourage changes in students both in attitudes and in science. With the learning process, the delivery of educational teaching materials can be achieved at the right target.
Based on the above background, it is necessary to have a method that can guarantee the learning process of schools, especially at the high school level. One method that can be done and applied is Online Learning. Online learning is a learning method that uses internet networks with accessibility, connectivity, flexibility, and the ability to generate various types of learning interactions. The research stated by Zhang et al., (2004) shows that the use of the internet and multimedia technology is able to change the way of conveying knowledge and can be an alternative to learning carried out in traditional classrooms. [4]. At the implementation level, online learning requires the support of mobile devices such as smartphones or cellphones, laptops, computers, tablets, and iPhones which can be used to access information anytime and anywhere. [5] Learning is not only fixated on conventional learning systems, but there is already a learning system that uses media. Learning using this media, such as a website, is an effort to take advantage of the development of information technology. Online learning media is infrastructure used in learning activities that is used as a communication tool, computer devices and internet networks as learning media. [3] Online learning media is a means used to support teaching and learning activities in the form of components such as computers, information systems and internet networks that allow students to do distance learning. [2]. Online learning media is a tool used to assist learning and teaching activities online or online so that the interaction process can be channeled into face-to-face, effective and efficient teaching.
Based on the definition, it can be concluded that online learning media is a transition to teaching and learning activities that are usually carried out conventionally by switching to using a medium, namely a website which is an effort to utilize the development of information technology.
The use of web-based technology has a major contribution to the world of education, including the achievement of distance learning goals. Various media can also be used to support the implementation of online learning. For example, virtual classes use Google Classroom, Edmodo, and Schoology services [6][7] [8], and instant messaging applications such as WhatsApp. [9]. Online learning can even be done through social media such as Facebook and Instagram. [10]. Online learning connects students with learning resources (databases, experts / instructors, libraries) that are physically separated or even far apart but can communicate with each other, interact or collaborate (directly / synchronously and indirectly / asynchronously). Online learning is a form of distance learning that utilizes telecommunications and information technology, for example the internet, CD-ROOM. [11].
The teaching and learning process is basically a process of communication and knowledge transfer in the form of delivering messages from sender to receiver. In the process of delivering the message, a medium is needed so that the message can be received properly. Media or facilities are a very important component in a communication process. The level of effectiveness of the learning media used is very influential on the extent to which a communication role will be accepted by the recipient quickly and precisely or vice versa. [12]. The very rapid development of information and communication technology (ICT) has influenced various aspects of human life, including the interaction between teachers and students. These interactions require the support of innovative, creative, precise, and effective instructional media. [13].
Previous research on Web-based learning (WBL) was carried out by Taruna Nasution in 2015. The research concluded that "Web base learning is learning that requires technological tools, especially information technology such as computers and internet access. In practice, web-based learning utilizes internet facilities as a medium for delivering learning information (materials) such as websites, e-mails, mailing lists, and news groups. " [14].
Next is the research conducted by Boy Indrayana and Ali Sadikin in 2020. The research reports that (1) Students are interested in the application of e-learning. (2) the application of e-learning makes it easier about students to attend lectures. (3) The application of e-learning can prevent transmission of covid-19 on campus. The application of e-learning has obstacles to students that live in remote areas. The application of e-learning makes students become independent and have the courage to express their opinions. [15] The problem of this study is the use of conventional learning methods, while having to follow government regulations regarding the implementation of learning during this pandemic. This study aims to design a web-based learning media (Webbased learning). This research is limited to the design of web-based learning media (Web-based learning) at Wiyata Mandala High School.
II. METHODS
Research with the research sample is Wiyata Mandala Senior High School Balai Batang Tarang, Sanggau Regency, West Kalimantan using the Prototype Model Software Development method and data collection techniques in the form of observation, interviews and literature study.
Model of Software Development Model Prototype
The method used in this software development uses a prototype model or which is a software development method in the form of a circular cycle, where there are staging that show each activity in carrying out system development in the design being carried out.
The prototype model is one of the models of software development methods that are suitable for designing an information system, because this prototype model functions to explore customer requirements in more detail but has a high risk of increasing project costs and time [16]. The stages of the prototype model [16] are described as follows: 1. Listening to needs At this stage the authors conducted the first stage of the prototype model, where data was collected by listening to customers in order to obtain the data needed. The customer referred to here be the object of research at SMAS Wiyata Mandala Balai Batang Tarang. Data collection was carried out by means of observation and direct interviews. Regarding how the conventional learning system is and what are the obstacles between students and teachers in the process of implementing conventional learning systems and finding out what kind of learning system they need. Broadly speaking, it is in this analysis process that the author has got a picture of the learning system that will be designed, such as features and functions of the software.
Build or improve a mock-up
At this design stage the writer identifies the design for the program display, interface and algorithm on the system to be designed. This is based on the problem in question and looks for a solution like what the customer expects. With functional goals from both students, teachers and admins. This design is made using UML symbol diagrams from use case diagrams, activity diagrams, sequence diagrams, class diagrams, ERD and LRS. [17][18].
View or test mock-ups
This stage is the last stage of the prototype model. This stage aims to get responses, responses and evaluations from SMAS Wiyata Mandala Balai Batang Tarang. The author only makes a system design, if SMAS Wiyata Mandala Balai Batang Tarang gives a positive response or conforms to the system being designed, then it can be continued at the stage of making the system. If the design of this system cannot satisfy the Wiyata Mandala Senior High School of Balai Batang Tarang, the design of this system will be evaluated and tested until specifications are found in accordance with the wishes.
Data collection technique
Data collection techniques in this study were carried out to discuss problems by explaining, describing, explaining, and interpreting or writing a situation or event which was then analyzed so that a conclusion or subject matter was discussed or to obtain a fact of the state of the learning method in SMAS. Wiyata Mandala Balai Batang Tarang. The following data collection techniques are used:
Observation
Observation is a direct data collection technique carried out at the research site by researchers on existing objects and subjects in order to find existing problems and find out the conditions of the school. The author directly observes the learning activities carried out by SMAS Wiyata Mandala Balai Batang Tarang. From these observations, the writer immediately noted the problems that were found during the observation, systematically without any questions and communication that had to involve other people at Wiyata Mandala High School, Balai Batang Tarang.
Interview
The technique of collecting data by interview is a question and answer activity from two directions. To get data and information from respondents about an existing learning system problem that can also be supporting data from observation activities, in order to get complete and accurate information. Interviews were conducted by the author of the teachers, staff and students at Wiyata Mandala Senior High School, Balai Batang Tarang to obtain information on problems that occurred in these learning activities.
Literature studies
It is a technique of collecting data by the author by quoting some of the readings related to the web-based learning system at SMAS Wiyata Mandala Balai Batang Tarang. Reference sources in the form of literature, books, journals and reference sources from the internet exist to provide a strong theoretical basis.
Needs Analysis
Needs analysis is a stage of identifying the functional requirements of the program associated with the proposed activity process. [19]. The analysis of the needs for the design of an information system for web-based learning learning media is as follows: A. User Requirements In the design of the information system for web-based learning learning media, there are three (3) levels of users in operating web-based learning learning media, namely: Needs Scenario Administration Section, Scenarios for Teacher's Section Needs, Student's Part Needs Scenario. B. System Requirements 1. Users must log in first to be able to access the application by entering a username and password so that the privacy of each user is maintained. 2. Users must log out when finished using the application. 3. The system can display and process data processing in accordance with the available menu. 4. The system can display data on the value of each meeting when students have finished working on the assignments for each meeting. 5. The system prints reports according to the search criteria
Use Case Diagram Design
The analysis of user requirements that has been described in the previous subchapter will be the basis of designing the system. The needs analysis will be modeled on a use case diagram. The results of modeling the user needs analysis into a use case diagram for the design of an information system for web-based education learning media for SMAS Wiyata Mandala Balai Batang Tarang can be seen in the image of:
Activity Diagram Design
Activity diagram is a description of the activities for each use case as shown in Figure 1. Activity diagrams in this study consist of: Activity diagram logins, Activity diagrams managing TU data, Activity diagrams manage department data, Activity diagrams manage subject data, Activity diagrams manage class data, Activity diagrams manage student data, Activity diagrams manage teacher data,Activity diagrams manage teaching teacher data, Activity diagrams manage announcements, Activity diagrams access student achievement reports, Activity diagrams manage teaching material data, Activity diagrams manage question bank data, Activity diagrams make homework, Activity diagrams manage meeting schedules, Activity diagram settings, Activity diagram download module, Activity diagrams do the problems, Activity diagrams of doing homework, Chat activity diagram, Activity diagram logout.
Here the authors describe some samples of the results of modeling activity diagrams for the design of a web-based learning media information system for webbased education at SMAS Wiyata Mandala Balai Batang Tarang which can be seen in the following figure.
Activity diagrams login
The business activities of the use case of Figure 1. will be modeled into an activity diagram. The results of modeling business activities from the login use case can be seen in the following figure :
Fig 2. Activity Diagram Login
The picture above explains how to enter the web-based learning learning media information system. First, the user must enter a username and password, then it is validated by the system. If successful, the user can access the main page according to the predetermined access level.
Prototype Design
At this stage, it will be explained about the interface design of web-based learning media at Wiyata Mandala Balai Batang Tarang High School which consists of three (3) access levels, namely Administration, Teachers and Students. The results of the interface design can be seen in the following figure.
A. Design of Administration Access Level Prototype
Administration is a user of the design of web-based learning media at Wiyata Mandala Senior High School Balai Batang Tarang. The results of the interface design for the administrative level are: 1. Administration login display; 2. Administration dashboard display; 3. Display TU data; 4. Display add data TU; 5. Display student data; 6. Display added student data; 7. Data display of majors; 8. Display added data majors; 9. Display of subject data; 10. Display added subject data; 11. Display class data; 12. Display added class data; 13. Display teacher data; 14. Display added teacher data; 15. Display teaching teacher data; 16. Display added teaching teacher data; 17. Display announcements; 18. Display added announcement data; 19. Display student achievement reports.
The following is the author describes some sample displays at the Administration level:
Display login administration
The login page is a user validation checking page. Administration must fill in a username and password in order to access web-based learning. The interface design for the login page (administration) can be seen in the following figure:
Administration dashboard display
The back-end (administrator) special dashboard can be accessed by the Administration if it has passed a series of validations at the login stage. This page is the main menu that provides sub menu processing, which consists of master data (TU data, students, majors, subjects, classes and teachers), managing (teaching teachers and announcements) accessing student achievement reports and logging out. The interface design for the dashboard (administration) can be seen in the following figure:
B. Design of Teacher Access Level Prototype
The teacher is a user of the web-based learning media design at Wiyata Mandala Senior High School Balai Batang Tarang. The results of the interface design for the teacher level are as follows: 1. Teacher login display; 2. Teacher dashboard display; 3. Display teaching material data; 4. Display added teaching material data; 5. Display question bank data; 6. Display add question bank data; 7. Display PR data; 8. Display added PR data; 9. Display meeting schedule data; 10. Display added meeting schedule data; 11. Chat display; 12. Display student achievement reports.
The following is the author describes some sample displays at the Teacher level: Teacher login view.
The login page is a user validation checking page. Teachers must fill in a username and password in order to access web-based learning. The interface design for the login page (teacher) can be seen in the following figure: The following is the author describes some sample displays at the Student level:
Display student login
The student login page is a page that functions as user validation, especially students. Students are required to fill in a username and password in order to access the web-based learning dashboard page.
Software Design
The software design is made using the UML method, including in the form of entity relationship diagrams (ERD), logical record structures (LRS), file specifications, class diagrams, sequence diagrams, and hardware and software specifications.
Entity Relationship Diagram (ERD)
Database design modeling using entity relationship diagrams (ERD) aims to describe the entities, attributes and relationships contained in the web-based learning media information system database at SMAS Wiyata Mandala Balai Batang Tarang. The results of entity relationship diagram (ERD) database design modeling can be seen in the following figure:
Logical Record Structure (LRS)
Based on the Entity relationship diagram (ERD) above, a logical record structure (LRS) can be described which aims to provide a clearer picture of the entities, attributes and relationships that are formed on a web-based learning database. The results of database modeling using a logical record structure (LRS) for the database on the design of a web-based learning media information system at SMAS Wiyata Mandala Balai Batang Tarang can be seen in the following figure. To determine the effectiveness and usefulness of the web-based learning media that has been made in this study, a survey was conducted on both students and teachers at SMAS Wiyata Mandala Balai Batang Tarang. The survey is created using the Google Form application and is given after the user tries to use the media as a whole. This survey aims to determine the level of media acceptance by students and teachers. The following are the items of the survey questions that have been distributed: 1. Does this web media help in the delivery or acceptance of lessons? 2. Be the teaching material presented in accordance with the teaching material? 3. Be the navigation of the web media displayed easy to understand? 4. Be the web media response speed good? 5. Be the information conveyed by the web media clear? 6. Can the use of web media save time? 7. Be web media easy to use? 8. Be the appearance and aesthetics of this web media attractive? 9. Be the home page display on this web media attractive? 10. Be the overall web media good?
IV. CONCLUSION
The results of the analysis that has been carried out based on the web-based learning media of SMAS Wiyata Mandala Balai Batang Tarang which have been designed and made in the form of a prototype, it can be concluded that: this web-based learning media for SMAS Wiyata Mandala Balai Batang Tarang provides facilities that are very helpful for administrators in managing administrative data, majors, subjects, classes, students, teachers, teaching teachers, announcements and access student achievement reports. From the teacher's point of view, the web-based learning media at SMAS Wiyata Mandala Balai Batang Tarang provides facilities that really help teachers to manage teaching materials, question banks, schedule meetings to make homework, chat and access student achievement reports. And in terms of students, the web-based learning media for SMAS Wiyata Mandala Balai Batang Tarang provides a very easy and interesting means of following the learning process such as accessing teaching materials or modules, doing homework, chatting, working on weekly meeting questions according to the provisions made by the teacher. and access the value of the outcome of the meeting. | 5,104.4 | 2021-05-05T00:00:00.000 | [
"Computer Science",
"Education"
] |
Motional consensus of self-propelled particles
The motional consensus of self-propelled particles is studied in both noise-free cases and cases with noise by the standard Vicsek model. In the absence of noise, we propose a simple method, using grid-based technique and defining the normalized variance of the ratio of the number of particles locally to globally, to quantitatively study the movement pattern of the system by the spatial distribution of the particles and the degree of aggregation of particles. It is found that the weaker correlation of velocity leads to larger degree of aggregation of the particles. In the cases with noise, we quantify the competition between velocity alignment and noise by considering the difference of the variety of order parameter result from the velocity alignment and noise. The variation of the effect of noise on motional consensus is non-monotonic for the change of the probability distribution of noise from uniform to non-uniform. Our results may be useful and encourage further efforts in exploring the basic principles of collective motion.
Model
Here, we consider the standard Vicsek model 17 including N self-propelled particles. All of the particles are regarded as points and continuously move in a two-dimensional square cell. The linear size of the square cell is L and the cell is considered with periodic boundary conditions. The interaction radius among particles is r which means the field of vision for each particle is πr 2 . Particles in the field of vision of particle i, including particle i itself, are regraded as the neighbor of it. The motion of each particle obey the ordinary differential equations (ODEs) as follows where N i (t) = {j : |x j − x i | r} denotes the number of the neighbor of particle i at time t and ξ θ i is random noise. And the ODEs of the model in 2D cartesian coordinates are as follows In the simulations, the time step updating the velocity and position of all of the particles is t = 1 . The initial positions of all of the particles are randomly distributed in the cell. Using the Euler method to discrete the Eq.(1), simultaneous update of the position of all of the particles at each time step according to and i denoting different particle among all of the N particles, takes one to N. All of the particles have the same absolute value of velocity v. The initial direction of velocity of all of the particles are randomly and uniformly distributed in [−π , π) . The variation of the velocity is shown in the change of the direction of the velocity. According to Eq.(1), the rule for updating the direction of velocity of the particles is where θ i (t) r is the average direction of all of the neighbors of the particle i, which is given by and �θ represents noise which is a random number following uniform distribution in [−η/2, η/2] . η is the parameter to control the amount of noise. In order to measure the degree of the motional consensus of the system, the normalized average velocity of all the N particles is considered as the order parameter of the system, which is as follows When the direction of the velocity of the particles achieves global alignment (flocking state), normalized average velocity φ reaches 1 and zero in the randomly disordered states 17 .
Result and discussion
The noise-free cases. In noise-free cases, η = 0 . When the number of simulated time steps is large enough, the order parameter of the system can reach 1. This means that the system arrives in a flocking state (strict motional consensus). In order to catch the main feature of motional consensus without sacrificing long simulation times, we take φ m = 0.979 to be the standard for reaching motional consensus. φ m = 0.979 , which means the system almost reaches flocking state, is large enough to ensure the validity of analysis of motional consensus below will not change. In simulation, we set L = 10 . The total time steps for simulation are 1000, which is long enough for the system to reach motional consensus.
Considering the rules of velocity alignment in the standard Vicsek model, the update of the direction of velocity will be directly influenced by their neighbors. Particles with common neighbors will build correlation of their velocity, while there will be no correlation of the velocity among the particles without common neighbor. To quantify simply, here, we only consider the correlation of each pair of particles consisting of two particles.
When the system reaches motional consensus, they are found to have different movement patterns which show different spatial distribution and different degree of aggregation of all of the particles. As shown in Fig. 1, with the increasing of the interaction radius r, the spatial distribution of the particles is more uniform and the particles aggregate less closer. As the total number of particles increases, the particles are distributed more evenly and clustered closer together. The increase in velocity makes the spatial distribution of particles more uneven and the particles aggregate more closer.
To quantitatively study the movement pattern of the systems reaching motional consensus, we proposed a method with grid-based technique. Grid-based technique is widely used in many numerical approaches including 45 and Multi-Particles Collision(MPC) etc 46 . FMM divides the space into different number of cells depending on the level of division. By analyzing the spatial relationship between the target cells and other cells, the interaction among particles in the target cell and the interaction of that with other cells can be obtained. Then the total interaction among the particles of the system can be evaluated. MPC introduces randomly shifted cells in the simulation of each time step. Evaluating the streaming and collision of the particles in each cell to obtain the position and velocity of the mass of the center of each cell. Then analyzing the dynamics of the system by considering the interaction among all of the cells. Both of the method mentioned above can improve the efficient of the simulation. But they are more suitable to solve the interaction among the particles. Here, we aim to catch the feature of the movement pattern of the system and our method with grid-based technique is simple and efficient. As shown in Fig. 2a, we divide the two-dimensional L × L space into G grids, where G = 25 here. Then we investigate the normalized variance of the ratio of the number of particles in each grid N i to total number of the particles N.
The rationality of current grid selection is discussed in the Supplementary Information. The ratio of the number of particles in each grid N i to total number of the particles N is obtained as follows The normalized variance of the ratio of N i to N is and σ is the variance of the ratio of N i to N where R = i R i /G = 1/G is the average of R i . σ max denotes the maximum of the variance of the ratio of N i to N when all of the particles aggreagate in the same grid, which is www.nature.com/scientificreports/ The normalized variance χ ratio will be 1 when all of the particles aggregate in the same grid, while χ ratio = 0 when all of the particles are uniformly distributed in G grid. As Fig. 2b-d shown, the increase of r or N and the decrease of v leads to smaller value χ ratio which means the more uniform spatial distribution of the particles ( Supplementary Information 1).
To quantify the degree of aggregation of the particles, we investigate the average number of particles for the grid that occupied by particles N grid . As Fig. 3a shown, with the increasing of r, the value of N grid becomes www.nature.com/scientificreports/ smaller, which means the aggregation of the particles is less close. As shown in Fig. 3b,c, N grid increases as N or v increase. The increase of N or v makes the particles aggregate more closer. In order to understand the variation of the degree of aggregation of particles and the spatial distribution of the particles with different r and N, we study the common neighbors of particles. The common neighbors of particles are significant for the movement pattern of the particles by affecting the update of the velocity and position of particles. For noise-free cases, the movement pattern of particles when reaching motional consensus depends only on the initial state of all of the particles. Given the initial state, the state of the system is fixed after updating in each time step, because the update in each time step is not affected by noise.
For the initial state, we first pay attention to the average ratio of the number of common neighbor n com to the number of particles within the field of vision of pairs of particles n pair = n A + n B − n com , where n A and n B are the number of neighbors of particle A and B, which are any two of the N particles. n com /n pair reveals the average strength of the correlation of velocity for each pair of particles. As Fig. 4a,b shown, the decrease of r or the increase of N leads to weaker correlation of velocity for each pair of particles.
Because the motional consensus of the system is not only affected by the pairwise correlations of the velocity, but also related to all of the particles, we also investigate the average ratio of the number of common neighbor n com to the total number of the particles N in initial state. As shown in Fig. 4c,d, with the decreasing of r or the increasing of N there is weaker correlation of velocity between pairwise particles and all of the particles.
What have been analyzed about Fig. 4 shows that increasing N and decreasing r will weaken correlation of velocity among particles. Weaker correlation of velocity among particles makes it more difficult for particles to reach motional consensus. The particles will keep moving in their direction respectively until the correlation between their velocity is large enough to enable them to move in the almost same direction. In order to build stronger correlation of their velocity, particles will move more closer, resulting in larger degree of aggregation when the system reaches motional consensus.
There is a different mechanism for the effect of velocity to the degree of aggregation of particles. For small velocity, particles move slowly to close in order to reach motional consensus by building stronger correlation of their velocity. Because particles move slowly, they are more sensitive to the boundary that whether they can reach www.nature.com/scientificreports/ motional consensus or not. With the increasing velocity, particles move fast and are insensitive to the boundary that whether they can reach motional consensus or not, which leads to larger aggregation of the particles.
Cases with noise. The rules of velocity alignment make the velocity of all of the particles unified, while the noise disturbs the motional consensus. For the motion of the particles, the restriction of velocity alignment will be weakened by the effect of noise, which makes the motional consensus more difficult or even impossible to reach. The temporal evolution of the order parameter for various strengths of noise are shown in Fig. 5a. There is perturbation of order parameter when the state of the system can be thought to be steady. In order to quantify the value of order parameter when the system is nearly steady, we take the average of order parameter from 500 steps to 1000 steps as the order parameter of the system in the nearly steady state. With the increasing of η , the steady value of the order parameter is smaller, which means the less unified of the motion of all of the particles. Figure 5b shows the order parameter for different values of η in the absence of velocity alignment. It is impossible to reach motional consensus even in the case of small η.
In order to quantify the effect of noise on motional consensus, we investigate the difference of order parameter �φ between two cases as follows where φ va denotes the steady value of order parameter in the cases that the motion of particles is only restricted by velocity alignment and φ vn denotes the steady value of order parameter in the cases that the motion is affected by both velocity alignment and noise.
As shown in Fig. 6a, the difference of order parameter �φ increases with the increasing of the value of η for different interaction radius r, which shows the larger effect of noise on motional consensus.
To compare the influence of velocity alignment and noise on reaching motional consensus and quantify the competition between the effects of velocity alignment and noise, we defined the difference of the difference of the order parameter κ which is shown as follows where �ϕ = φ va − φ n denotes the difference of order parameter between the cases that the update of the motion of particles is affected by both velocity alignment and noise and the cases that the motion is just affected by noise. As shown in Fig. 6b, the value of κ will change from positive to negative as η increases, which means the increasing of η improves the effect of the noise to the motion of the particles and the effect of velocity alignment becomes more and more weak comparing that with the noise. When κ = 0 , velocity alignment and noise affect motional consensus equally.
We also observed that the variation of �φ is not monotonous when the value of η is larger than 6. When η = 6 , �θ ∈ [−3.0, 3.0] , which is close to [−π, π] , the probability of all the direction of velocity effected by noise is almost equal as shown in Fig. 7a.
In order to confirm our analysis of the reason of the non-monotonic variation of �φ when η > 6 , we investigate the probability distribution of noise with different value of η . As shown in Fig. 8, different values of η lead to different probability distribution of �θ , which affect the value of �φ. This is consistent with the above analysis of noise. As shown in the inset of Fig. 7a, the degree of the nonmonotonic variation about the difference of order parameter is larger with the increasing of the interaction radius r. This is because the larger interaction radius improves the effect of velocity alignment on motional consensus, The red area and the green area denotes the probability when �θ ∈ [−1.2π, π] and �θ ∈ [π, 1.2π] respectively before unifying and they denote �θ ∈ [−π, −0.8π] and �θ ∈ [0.8π, π] after unifying.
Conclusion
In conclusion, we have studied the motional consensus of self-propelled particles in both the noise-free cases and the cases with noise by standard Vicsek model. For the noise-free cases, we have proposed a method to quantitatively describe the spatial distribution of the particles by divided the two-dimensional space into some grid with equal size and count the normalized variance of the ratio of N i to N. It is found that the smaller r or larger N builds weaker correlation of the velocity among particles, which leads to larger degree of aggregation of the particles when the system reaches motional consensus.
For the cases with noise, we have quantitatively analyzed the competition between the effects of velocity alignment and noise on the degree of motional consensus. The results show that the non-monotonic variation of the effect of noise on motional consensus result from the non-uniform probability distribution of the noise.
Collective behaviors of active systems present various patterns. Pattern formation of active systems may be studied by generalizing the Smoluchowski aggregation theory which focus on the growth and distribution of clusters for passive systems 47 . Bridging the Vicsek model(particles-based model) and the theroy proposed by Tu and Toner based on hydrodynamics 48 is also an interesting perspective of study about collective behavior of active systems. As for the further studies concerning the collective behavior of active systems, our study may be useful for exploring the basic principle of collective motion.
Data availability
The data presented in this study are available on request from the corresponding author. | 3,837.4 | 2023-05-20T00:00:00.000 | [
"Physics"
] |
Adaptive spatial filtering of daytime sky noise in a satellite quantum key distribution downlink receiver
Abstract. Spatial filtering is an important technique for reducing sky background noise in a satellite quantum key distribution downlink receiver. Atmospheric turbulence limits the extent to which spatial filtering can reduce sky noise without introducing signal losses. Using atmospheric propagation and compensation simulations, the potential benefit of adaptive optics (AO) to secure key generation (SKG) is quantified. Simulations are performed assuming optical propagation from a low-Earth-orbit satellite to a terrestrial receiver that includes AO. Higher-order AO correction is modeled assuming a Shack–Hartmann wavefront sensor and a continuous-face-sheet deformable mirror. The effects of atmospheric turbulence, tracking, and higher-order AO on the photon capture efficiency are simulated using statistical representations of turbulence and a time-domain wave-optics hardware emulator. SKG rates are calculated for a decoy-state protocol as a function of the receiver field of view for various strengths of turbulence, sky radiances, and pointing angles. The results show that at fields of view smaller than those discussed by others, AO technologies can enhance SKG rates in daylight and enable SKG where it would otherwise be prohibited as a consequence of background optical noise and signal loss due to propagation and turbulence effects.
Introduction
The threat quantum computing poses to public key cryptography is motivating the development of alternatives to modern key sharing techniques that rely on computational complexity for security. 1,2 Presently, there is interest in developing quantum key distribution (QKD), presented by Bennett and Brassard in 1984 (BB84), as a provably secure alternative. [2][3][4][5] QKD lends itself to mathematical proofs of theoretical security and offers the potential for secure generation of symmetric encryption keys in real time over optical channels. 6,7 The BB84 QKD protocol generates encryption keys using polarization states of light transmitted and detected via individual photons. Attempts by an eavesdropper to intercept, clone, and resend individual photons lead to errors in the cloned states that in turn lead to bit errors, referred to as quantum bit errors. 3,8 Quantum bit errors can be detected by the key sharing parties to reveal the presence of the eavesdropper. Based on the assumption that a technologically advanced eavesdropper could suppress naturally occurring bit errors, all bit errors are assumed to be due to eavesdropping and an indication of information leakage. This includes bit errors that may in actuality be due to scattered light, detector dark counts, and channel crosstalk. The inferred information leakage is mitigated with privacy amplification algorithms that reduce the number of key bits. 9 For sufficiently large quantum bit error rates (QBERs), secure key generation (SKG) is not possible. 10 It is therefore important to consider technologies that minimize naturally occurring sources of bit errors.
Concepts for global QKD networks include the use of free-space quantum channels linking ground-and spacebased nodes. [11][12][13] Demonstrations of free-space QKD have been carried out successfully, including implementations over terrestrial 12,[14][15][16][17][18] and ground-air quantum channels. 18,19 Free-space quantum channels present a number of practical challenges for QKD. The scattering of ambient light into the quantum channel can be a significant source of quantum bit errors. Signal transmission efficiencies are reduced by beam divergence, atmospheric scattering and absorption, and atmospheric-turbulence-induced wavefront errors. The mechanisms of noise and loss both contribute to increased QBERs that can preclude SKG. Terrestrial demonstrations of free-space QKD in daylight have utilized spectral and temporal filtering techniques to reduce noise due to scattered light. This includes a demonstration conducted at a low daytime sky radiance using a dielectric spectral filter 14 and a demonstration conducted at significantly higher radiance using an atomic-line spectral filter. 15 System-level analyses for implementations linking low-Earth orbit (LEO) satellites to terrestrial ground stations support plans for full-scale satellite demonstrations. 18,[20][21][22][23][24][25][26][27] Studies have considered both upward and downward propagating signal photons and concluded downlinks offer the advantage of lower transmitter-to-receiver aperture coupling losses due to the effects of atmospheric turbulence. 13,22 In the downlink scenario, the receiver points toward the sky.
In this scenario, sky radiance can be a significant source of quantum bit errors in daytime. Analyses of daytime satellite QKD downlinks have assumed a variety of sky radiance values under cloud-free conditions. 20,22,28 Recently, we presented an analysis of hemispherical distributions of daytime sky radiance and QBERs associated with satellite transmitters in circular orbits under clear sky conditions. 29 Spatial filtering can also mitigate sky noise in a QKD downlink receiver. This is accomplished by reducing the receiver field of view (FOV). Atmospheric turbulence limits the extent to which this can be done without introducing signal loss. FOVs previously discussed in the literature are sufficiently large to avoid turbulence-induced signal loss. [14][15][16]18,22,24 It has been discussed that improved tracking, or wavefront tilt correction, in a QKD receiver can improve the signal at small FOVs. 17,20,30,31 Recently, we presented preliminary results from numerical simulations showing that implementing higher-order adaptive optics (AO) at reduced FOVs in a QKD ground-station receiver can significantly improve SKG rates in daytime and enable SKG under conditions where it would otherwise be prohibited. 32 The concept of real-time sensing and correction of atmospheric-turbulence-induced wavefront errors was proposed in 1953 by Babcock. 33 The first practical design for atmospheric compensation was proposed, patented, and demonstrated in the mid-1970s by Hardy while working for Itek optical systems. [34][35][36] Compensated imaging of LEO satellites was first accomplished in 1982 by the Air Force under funding from the Advanced Research Projects Agency. 36 This demonstration was followed by the rapid development of supporting and enabling technologies. [36][37][38] In modern AO systems, low-order and higher-order wavefront errors are treated separately. Low-order errors corresponding to linear wavefront tilts are compensated with steering mirrors. Higher-order errors are compensated using two-dimensional wavefront compensating optics. Optimized compensation of dynamic wavefront errors requires closed-loop control of the AO system at a bandwidth that typically exceeds the rate at which turbulence is changing. 39 Where telescope slewing is involved, the wavefront error includes a translational component associated with both wind and slewing. This translational component is described by the Greenwood frequency. 39 High Greenwood frequencies associated with tracking a LEO satellite through turbulence can challenge the performance capabilities of an AO system.
Atmospheric turbulence is a stochastic process. Consequently, propagation through turbulence leads to statistical distributions of wavefront errors. In a free-space quantum channel, this can lead to statistical distributions of SKG rates. Quantifying the effects of turbulence and AO on SKG rates requires numerical simulations. The Air Force Research Laboratory sponsored the development of fully integrated software that accurately models both the effects of wave propagation through the atmosphere and AO compensation. 40,41 The software propagates optical wavefronts through statistical representations of turbulence and through the optical components of an AO system including the time-dependent behavior of these components under closed-loop control.
This paper considers a LEO satellite QKD downlink to a terrestrial receiver that includes an AO system. Numerical simulations quantify the effects of turbulence, FOV, and AO on photon capture efficiency in the receiver. The simulations consider moderate and strong turbulence with elevation angles ranging from zenith to 15 deg above the horizon and moderate and high daytime sky radiance for satellites in 400-and 800-km altitude circular orbits. The simulations include the pointing-angle-dependent optical losses associated with optical scattering and absorption, transmitter-toreceiver aperture coupling, and atmospheric turbulence. SKG rates are calculated for a decoy-state QKD protocol for the case of tilt correction only and compared to the case where higher-order wavefront correction is applied. Results show that in the presence of moderate turbulence and moderate sky radiance, simply reducing the receiver FOV can reduce sky noise sufficiently to enable SKG with tilt correction alone. In this case, the addition of higher-order AO enhances SKG rates considerably. Under more challenging conditions of strong turbulence, high sky radiance, and longer propagation distances, higher-order AO can enable SKG where it would otherwise be prohibited due to noise and loss.
The Effects of Turbulence on Signal and Noise
Transmission in an Optical Receiver The scattering of sunlight by the atmosphere into the quantum channel leads to spurious detection events and quantum bit errors. Atmospheric turbulence contributes to this problem by increasing the size of the signal distribution at the receiver field stop (FS). In order to avoid signal losses due to turbulence, the size of the FS is increased relative to that required by the diffraction limit. This results in increased levels of sky noise transmitted by the FS to the detectors.
Atmospheric Scattering-Induced Noise
Figure 1(a) shows the basic elements of an optical receiver including a primary optic of diameter D R that defines the entrance pupil, an FS of diameter d, a collimating lens, and a spectral filter with bandpass Δλ. The FS limits the chief ray of the system to define the linear angle FOV, Δθ ≈ d∕f, where f is the focal length of the primary optic. Reducing the size of the FS reduces the FOV and, correspondingly, the number of sky noise photons transmitted by the FS.
The number of sky-noise photons, N b , entering the receiver in a detection window is proportional to the sky radiance according to the radiometric expression 20 E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 1 ; 3 2 6 ; 2 5 8 where H b is the sky radiance in W∕ðm 2 sr μmÞ, Ω FOV ¼ πΔθ 2 ∕4 is the receiver solid-angle FOV, R is the radial extent of the primary optic, λ is the optical wavelength, Δλ is the spectral filter bandpass in μm, Δt is the integration time for photon counting, h is Planck's constant, and c is the speed of light. For a given value of the sky radiance, the number of noise photons may be reduced by reducing the spectral bandpass, the temporal gate width, and the receiver FOV.
Atmospheric-Turbulence-Induced Aberrations
Figure 1(b) shows the effects of turbulence on signal transmission at the FS. In the absence of aberrations, the primary optic focuses an incident signal to a diffraction-limited irradiance distribution. For a signal photon derived from an attenuated laser pulse, the irradiance distribution represents the photon probability function. For the case of a planar wavefront with uniform amplitude incident upon a circular aperture, the diffraction-limited irradiance distribution at focus is described by the Airy function with a central disk of diameter 2.44λf∕D R . Reducing the FS to this diameter passes 84% of the signal while blocking sky noise associated with larger field angles. Reducing the FS further decreases both the transmitted sky noise and the signal. Aberrations increase the size of the signal irradiance distribution at the FS. For atmospheric-turbulence-induced aberrations, the strength of turbulence integrated over a propagation path is characterized by Fried's coherence length r 0 . 42 Sarazin and Roddier 43,44 computed the angular FWHM of the long-exposure irradiance distribution with turbulence to be approximately λ∕r 0 . In analogy to the Airy disk, we define the turbulence-induced spot size to be that which captures about 84% of the power. For primary optic diameters larger than r 0 , this aberrated spot size at focus is found to be approximately 2λf∕r 0 . Relative to the diffraction-limited case, high transmission requires the diameter of the FS to be increased from 2.44λf∕D R to 2λf∕r 0 . This increases the linear FOV from Δθ ≈ 2.44λ∕D R to Δθ ≈ 2λ∕r 0 and, correspondingly, increases the number of noise photons transmitted by the FS by a factor of ðD R ∕1.22r 0 Þ 2 .
In principle, an AO system can restore the aberrated wavefront to near-diffraction-limited quality. Within the boundaries established by diffraction and turbulence, AO could play a significant role in preserving the transmission of signal photons at reduced FOVs that would substantially reduce background sky noise. The potential benefit of AO to sky noise reduction can be estimated as follows. Atmospheric turbulence is characterized by standard altitude-dependent turbulence profiles. The strength of turbulence is both angle-dependent and wavelength-dependent. For the commonly used HV 5∕7 turbulence profile, 45 henceforth referred to as 1 × HV 5∕7 , and a wavelength of 780 nm, r 0 ranges from about 9 to 4 cm for pointing angles ranging from 0 deg to 75 deg from zenith, respectively. For a D R ¼ 1 m diameter receiver aperture, the corresponding range of turbulence-limited FOVs is 18 to 40 μrad. In the absence of turbulence, the diffraction-limited FOV is ∼2 μrad. At 75 deg from zenith, reducing the FOV from the 40 μrad turbulence-limited FOV to the 2 μrad diffraction-limited FOV would reduce sky noise by a factor of 400. Assuming a perfect AO system, this reduction in optical noise could be achieved without increasing the signal loss at the FS.
Quantifying the Effects of Turbulence and
Adaptive Optics on Photon Capture Efficiency In practice, AO does not completely compensate for turbulence-induced aberrations. Limitations occur due to the finite spatial resolution and finite temporal response of the AO system. For a given set of atmospheric parameters and AO system specifications, the optical efficiency of the system can be calculated using numerical methods. In the analysis that follows, it is assumed that the transmission efficiencies associated with classical irradiance distributions represent the transmission probabilities for individual photons from an attenuated laser pulse. It is also assumed that the FS is imaged onto the photodetectors without loss of energy due to vignetting.
Quantum Key Distribution Receiver with
Adaptive Optics Figure 2 shows a conceptual schematic of an optical receiver that includes both a QKD receiver and an AO system. The AO system architecture is based on systems previously demonstrated. 36,37 AO systems require light from either a natural or artificial beacon to probe the atmospheric turbulence. 36,37,46 Control signals for the AO system are generated from measurements performed on the aberrated beacon wavefront. In the analysis that follows, it is assumed that the satellite includes both an AO beacon laser and a QKD photon source. The two sources generate copropagating optical pulses that are synchronized in time, but at different wavelengths to allow chromatic separation at the receiver. Wavefront errors caused by atmospheric turbulence consist of a tilt component that causes an image to jitter and higher-order spatial components that cause the point spread function to enlarge. Correspondingly, the AO system consists of a fast steering mirror (FSM) that tracks the tilt component of the wavefront error and a deformable mirror (DM) that compensates higher-order wavefront errors. A dichroic beam splitter (DBS) diverts the beacon wavefront to the wavefront sensing system. A beacon-channel spectral filter (BCSF) transmits the beacon wavelength while blocking other spectral components. A beam splitter (BS) transmits a portion of the beacon light to a focal plane tracking sensor (FPTS) that measures the tilt component of the wavefront error and generates control signals for the FSM. The reflected light propagates to a Shack-Hartmann wavefront sensor (SHWFS) that determines the higher-order aberrations and generates control signals for the DM. In a closedloop AO system, dynamic feedback control reduces residual wavefront errors. The AO system's dynamic performance limit is characterized by the system bandwidth. Beyond this frequency, the amplitude response of the system is insufficient to be considered useful or stable. The quantum channel wavelength is transmitted by the DBS, reflected by a mirror (M), and brought to a focus at the FS of the system. A quantum-channel spectral filter (QCSF) transmits the quantum-channel wavelength to the QKD receiver while blocking other spectral components. Within the QKD receiver, a 50/50 BS randomly directs photons to the two mutually unbiased measurement bases of the BB84 protocol. In the reflected path, a polarizing beam splitter (PBS) and two gated avalanche photodiodes (APDs) measure the polarization state of the photon in the rectilinear basis. In the transmitted path, a half-wave plate (HWP) rotates the polarization states by 45 deg for measurement in the diagonal polarization basis.
Numerical Simulation of Propagation Through
Turbulence and an Adaptive Optics System The photon capture efficiency associated with turbulence, η spatial , is defined to include the turbulence-related losses associated with both transmitter-to-receiver aperture coupling and propagation through the FS of the receiver. Aperture-toaperture coupling losses due to diffraction are accounted for separately. The term η spatial is quantified with a simulation code that includes the effects of turbulence on wave propagation and the effects of a closed-loop AO system within the receiver as shown in Fig. 2. Wave-optics simulations are performed with Atmospheric Compensation Simulation, a simulation code developed by Science Applications International Corporation. 40,41 The wave-optics propagation code has been anchored to experimental data and used to anchor other simulation codes. 47 The code is based on the principles of scalar diffraction theory. The simulated AO system includes an FPTS and FSM for tilt estimation and correction and an SHWFS and DM for higher-order aberration correction. The hardware emulations include models for the wavefront sensor cameras that include real world effects such as noise, pixel diffusion, and latencies. Simulations are performed for receiver pointing angles ranging from zenith to 75 deg from zenith. For each elevation angle, 20 realizations of atmospheric turbulence are simulated. For each realization of turbulence, the atmosphere is simulated by 10 phase screens distributed throughout the atmospheric path. Each phase screen is a random realization of turbulence consistent with Kolmogorov statistics and the specified turbulence strength profile. Numerical methods based on scalar Fresnel integrals 48,49 propagate the optical field from the transmitter through the phase screens 50 to the receiving aperture. From there, the optical field is reflected from an FSM and then a DM. These components are simulated in a closed-loop for iterative feedback control.
Adaptive Optics System Parameters
The FPTS is modeled as a lens and focal plane quadrant detector. The SHWFS is modeled as a 32 × 32 element array of lenslets and quadrant detectors. The SHWFS is assumed to be shot-noise limited as is typically the case for systems with cooperative beacons. The DM is modeled as a continuous face sheet driven by a 33 × 33 array of actuators. The diameter d l of the individual lenslets in pupil space is such that d l ∕r 0 is less than unity for the range of turbulence strengths encountered in a 1 × HV 5∕7 atmospheric profile. In stronger turbulence, where d l ∕r 0 exceeds unity, the system performance is degraded. 46 In the simulations, the FPTS centroid, SHWFS centroids, and residual wavefront errors are updated at 10 kHz. The FSM and DM are also updated at 10 kHz. The tracking bandwidth is 200 Hz and the bandwidth for higher-order correction is 500 Hz. These system parameters are considered to be within the state of the art.
It is assumed that the cooperative beacon on the satellite provides light at 810-nm wavelength for the FPTS and SHWFS. The quantum channel wavelength is assumed to be 780 nm, allowing separation of the two wavelengths. It is further assumed that any beacon light transmitted by the DBS and QCSF is insignificant compared to other noise sources. Applying wavefront correction at a wavelength that is shorter than the beacon wavelength can lead to residual wavefront errors. While these errors are accounted for in the simulations, they are negligible due to the small separation in wavelengths. Similarly, the beacon and quantum-channel pulses are separated in time, but on a timescale over which the atmosphere is static in the simulations.
For the purpose of analysis, it is assumed that the satellite travels in either a 400-or 800-km altitude circular orbit. The telescope slews to follow the motion of the satellite. The altitude-dependent wind speed is described by the Bufton wind profile. In order to consider the worst-case scenario, the wind direction is assumed to be opposite to the slew direction, producing the highest Greenwood frequency for a particular turbulence profile.
Atmospheric Turbulence Parameters
The effects of turbulence on an optical field are characterized by temporal, angular, and spatial coherence parameters. The temporal coherence, given by the Greenwood frequency f G , is dependent upon the slew rate of the telescope and, therefore, the altitude of the satellite. The angular coherence is given by the isoplanatic angle θ 0 . The spatial coherence is given by Fried's coherence length r 0 . Greenwood frequencies exceeding the correction bandwidth and isoplanatic angles smaller than the angular subtense of the source result in a degraded performance of the AO system. Rytov is a direct measure of scintillation experienced by the optical field at the receiver entrance pupil. Rytov values greater than about 0.4 indicate deep turbulence where scintillation leads to degradation in the performance of the AO system. In the presence of scintillation, the SHWFS is unable to accurately measure wavefront errors due to intensity nulls in the field. Table 1 shows turbulence parameters calculated at each of the five elevation angles for the two turbulence profiles and two orbit altitudes considered in the analysis that follows. For a 10-cm transmitter aperture, the isoplanatic angles are larger than the angular subtense of the source for all cases. For the 800-km orbit, the Greenwood frequency remains within the 500 Hz AO system bandwidth for all cases shown except 75 deg in 3 × HV 5∕7 turbulence. For the 400-km orbit, the higher slew rates cause the Greenwood frequency to exceed the AO system bandwidth at 75 deg from zenith in 1 × HV 5∕7 turbulence and at all angles shown below zenith in 3 × HV 5∕7 turbulence. At the 75 deg elevation angle, Rytov values indicate deep turbulence conditions. For D R ¼ 1 m, the SHWFS subaperture size in pupil space is d l ¼ 3.1 cm. In 3 × HV 5∕7 turbulence, r 0 is smaller than Table 1 Turbulence parameters for five elevation angles relative to zenith with 400-and 800-km altitude circular orbit altitudes. Parameters include Fried's coherence length r 0 , the isoplanatic angle θ 0 , Rytov, and the Greenwood frequency f G . Parameters are shown for 1 × HV 5∕7 and 3 × HV 5∕7 turbulence profiles.
Zenith angle (deg) 400 km altitude 800 km altitude the subaperture size at the 60 deg and 75 deg elevation angles leading to under-resolved local wavefront tilts.
Results: Photon Capture Efficiency Probability Distributions
The turbulence-related photon capture efficiency η spatial is calculated for each elevation angle from the 20 realizations of atmospheric turbulence. For each realization, the simulated AO system iterates to steady state minimizing the wavefront error at the SHWFS. Figure 3 shows sample results presented as probability density functions (PDFs) calculated for 1 × HV 5∕7 turbulence, a D R ¼ 1 m receiver aperture size, and a pointing angle at zenith. The transmitter is treated as an unresolved point source relative to a 16-cm grid spacing at the 400-km orbit. Capture efficiencies calculated with tilt correction only are shown in red. Capture efficiencies calculated with the addition of higher-order AO are shown in blue. Since tilt correction is required for tracking the satellite, the case without atmospheric tilt correction is not considered here.
Figures 3(a) and 3(b)
show results calculated for a 20-μrad FOV with 400-and 800-km altitude circular orbits, respectively. At this FOV, the mean capture efficiencies increase from 88% with tilt correction alone to 93% with the addition of higher-order AO. Figures 3(c) and 3(d) show the corresponding results calculated for a 2-μrad FOV. At the reduced FOV, turbulence leads to significant losses at the FS. The mean capture efficiencies now increase from only about 5.5% with tilt correction alone to about 73% with the addition of higher-order wavefront compensation. In both cases, the capture efficiency improves measurably with the addition of higher-order AO, but does not reach unity. This is due to the imperfect nature of the AO system and the fact that the small turbulence-related losses that occur in aperture-to-aperture coupling cannot be recovered by any AO system implemented in the receiver. The relative benefit of AO is more significant at the smaller FOV where losses due to turbulence are greater. Increasing the orbit altitude from 400 to 800 km decreases the Greenwood frequency. However, this has a negligible effect on the results since, for the examples shown, all turbulence parameters are within the correction tolerances of the AO system modeled.
Secure Key Generation Rates with a Decoy-State
Quantum Key Distribution Protocol The final measure of effectiveness for a QKD system is the rate at which key bits can be generated in a secure key sharing protocol. Methods for estimating SKG rates in the original BB84 protocol assume that potential information leakage to an eavesdropper can be inferred from the measured QBER and that a privacy amplification algorithm will be implemented to reduce information leakage to meet security criteria. 51 An information theoretic upper bound on information leakage determines the fraction of raw key bits to be retained in the secure key.
High-loss quantum channels are particularly sensitive to photon-number-splitting (PNS) attacks on multiphoton pulses. 52 In the PNS attack, the eavesdropper selectively blocks single-photon pulses where eavesdropping would introduce errors, and retains a photon from multiphoton pulses, allowing the eavesdropper to gain complete knowledge of the bit value without introducing errors. The decoy state protocol was proposed 53 and developed 54 to address this problem in high-loss channels. The protocol introduces decoy pulses with a mean photon number that is different from that of the signal pulses. Measurements of the decoy pulse detection yield allow the key sharing parties to detect the PNS attack and more accurately estimate possible information leakage from the measured QBER.
Secure Key Rate Equations for the Vacuum-Plus-Weak-Decoy-State Quantum Key Distribution Protocol
This section reviews the rate equations for the vacuumplus-weak-decoy-state QKD protocol implemented via polarization encoding of photons from a Poissonian light source as presented by Ma et al. 55 The calculation includes the effects of sky noise, detector dark counts, polarization crosstalk, mean photon number, and photon losses. The SKG rate per signal state, or secret bit yield, is given by E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 2 ; 6 3 ; 6 9 7 R ≥ qf−Q μ fðE μ ÞH 2 ðE μ Þ þ Q 1 ½1 − H 2 ðe 1 Þg; (2) where the protocol efficiency q is 1/2 for the BB84 protocol, μ is the mean photon number of the signal states, Q μ is the gain of the signal states, E μ is the overall QBER, Q 1 is the gain of the single-photon states, e 1 is the error rate of single photon states, fðE μ Þ is the bidirectional error correction efficiency, and H 2 is the Shannon binary entropy function. The gain of the signal states is given by E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 3 ; 6 3 ; where Y 0 is the background detection probability, 1 − e −ημ is the signal detection probability, and η is the efficiency of signal photon transmission and detection. The lower bound for the gain of the single-photon states is given by E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 4 ; 6 3 ; 5 1 0 Q 1 ¼ where ν denotes the mean photon number for the weak decoy state, ν < μ, and Q ν is the gain of the weak decoy state given by substituting ν for μ in Eq. (3). The upper bound of e 1 is given by E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 5 ; 6 3 ; 4 2 1 e 1 ¼ where Y 1 is the lower bound for the yield of the singlephoton states given by E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 6 ; 6 3 ; The overall QBER associated with signal photons is given by E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 7 ; 6 3 ; 2 8 where e 0 is the error rate due to noise and e detector is the probability that an incorrect bit value occurred due to polarization crosstalk. The background detection probability is calculated including contributions from sky radiance and detector dark counts E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 8 ; 6 3 ; 1 7 8 Y 0 ¼ N b η receiver η spectral η detector þ 4f dark Δt; (8) where N b η receiver η spectral η detector is the probability of detecting a sky-noise photon, η spectral is the efficiency of transmission through the spectral filter, η detector is the efficiency of photon detection, and η receiver is the efficiency of transmission through the remaining receiver optics. 14 The probability of a detection event occurring due to detector noise is 4f dark Δt, where f dark is the detector dark count rate at each of four identical detectors. The total signal transmission efficiency η also includes angle-dependent terms associated with propagation from the transmitter aperture through freespace, including the atmospheric path E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 9 ; 3 2 6 ; 7 0 8 η ¼ η geo η trans η spatial η receiver η spectral η detector ; (9) where η geo is the angle-dependent geometrical coupling efficiency between the transmitter and receiver apertures due to diffraction and finite aperture sizes, η trans is the angle-dependent transmission efficiency associated with atmospheric scattering and absorption, 14 and η spatial is the angle-dependent transmission efficiency associated with atmospheric turbulence.
Satellite-to-Earth Quantum Channel Parameters
SKG rates are calculated with the following parameters assumed: The quantum channel wavelength is λ ¼ 780 nm for low beam divergence, high atmospheric transmission, and high detector efficiency. 14 The beacon channel wavelength is 810 nm. The radial extent of the receiver primary optic is R ¼ 0.5 m. The spectral filter bandpass and transmission are Δλ ¼ 0.2 nm and η spectral ¼ 0.9, consistent with a commercially available dielectric interference filter. 56 The detector efficiency, dark count rate, and gate duration are η detector ¼ 0.6, f dark ¼ 250 Hz, and Δt ¼ 1 ns consistent with the operation of a commercially available Geigermode APD. 57 The collective optical efficiency of the remaining components in the receiver is η receiver ¼ 0.5. It is assumed that the output of the quantum channel source is expanded to uniformly illuminate the transmitter exit pupil and that all losses incurred by vignetting at the transmitter aperture contribute to the attenuation that is required to achieve the mean photon numbers, μ and ν. Under this assumption, it is not necessary to account for losses within the transmitter. For the aperture sizes and propagation ranges considered, the angle-dependent aperture-to-aperture coupling efficiency η geo can be approximated by the Friis equation, η geo ¼ ðπD T D R ∕4λzÞ 2 , which assumes a uniformly illuminated transmitter aperture. 58,59 The diameters of the transmitter and receiver apertures are D T ¼ 10 cm and D R ¼ 1 m, respectively. The propagation distance z is calculated as a function of the receiver pointing angle and satellite altitude. For a 400-km altitude circular orbit, propagation distances range from 400 km at zenith to 1175 km at 75 deg from zenith. For an 800-km orbit, the propagation distances range from 800 km at zenith to 2033 km at 75 deg from zenith.
The angle-dependent transmission efficiencies η trans are calculated with MODTRAN assuming clear sky conditions. Assumed values range from η trans ¼ 0.92 at zenith to η trans ¼ 0.74 at 75 deg from zenith. In the absence of turbulencerelated losses, η spatial ¼ 1 and the signal transmission efficiencies η, expressed in dB of loss, are ∼18 to 28 dB for the 400-km orbit and 24 to 33 dB for the 800-km orbit.
The background detection probability Y 0 is calculated from Eq. (8) with the number of noise photons N b calculated from Eq. (1). The gain of the signal state Q μ is calculated from Eq. (3) with the signal transmission efficiency η calculated from Eq. (9), including contributions from η geo , η trans , and η spatial . The signal QBER E μ is calculated from Eq. (7) with the error rate due to noise, e 0 , taken to be 1/2 under the assumption the background is random. It should be noted that the polarization distribution of sky radiance has been studied in detail, 60 but is not included in this analysis. The probability of a projective polarization measurement in a matched polarization basis yielding an incorrect result due to depolarization in propagation and improper alignment has been measured experimentally in a satellite-Earth optical link. 61 The following analysis assumes the reported measured value of e detector ¼ 2.8%. The gain of the single-photon states Q 1 is calculated from Eq. (4), assuming a decoy-state mean photon number ν ¼ 0.05. In order to specify an optimized mean photon number for the signal state, Eq. (2) was evaluated against μ as a free parameter. The mean photon number μ ¼ 0.45 was determined to be an optimum value for the parameters assumed here. 29 The single-photon error rate e 1 is calculated from Eq. (5) with Y 1 calculated from Eq. (6). The efficiency of error correction fðE μ Þ is assumed to be a constant value of 1.22, the commonly used value associated with Cascade error correction.
Results: Secure Key Generation Rate Probability Distributions
The statistical nature of turbulence leads to statistical distributions of SKG rates. SKG rates are calculated by evaluating Eq. (2) in Sec. 4.1 for statistical distributions of η spatial similar to those described in Sec. 3.5. Figures 4(a) and 4(b) show examples of SKG rate PDFs calculated from the capture efficiencies η spatial shown in Figs. 3(c) and 3(d), respectively, for the case of a 2 μrad FOV at zenith. This example assumes a 1 × HV 5∕7 turbulence profile and a sky radiance of 25 W∕ðm 2 sr μmÞ. In order to present SKG rates in units of secure key bits per second, a system rate of 10 MHz is assumed as a multiplicative factor to Eq. (2). Figure 4(a) shows results calculated for the 400-km altitude orbit. SKG rates with tilt correction alone are shown in red with a distribution around a mean value of 406 Hz. Note that a significant probability exists for a key rate of zero. Introducing higher-order AO results in a significant increase in SKG rates as shown in blue with a distribution around a mean value of 8.3 kHz. Figure 4(b) shows results calculated for the 800-km orbit. With tilt correction alone, the mean SKG rate is only 0.1 Hz, with zero being the most probable key rate. Introducing higher-order AO increases the mean SKG rate to 1.9 kHz. The SKG rates are significantly lower for the 800 km orbit as a consequence of the reduced geometrical capture efficiency η geo at the longer propagation distances.
Simulation Results: Daytime Key Rates with
Wavefront Tilt Correction and Higher-Order Adaptive Optics Numerical simulations are performed for turbulence conditions described by 1 × HV 5∕7 and 3 × HV 5∕7 turbulence profiles with receiver elevation angles ranging from zenith to 75 deg from zenith. Sky radiance values of 25 and 100 W∕ðm 2 sr μmÞ are considered with receiver FOVs ranging from 0.5 to 20 μrad. The case with tilt correction alone is compared to the case where higher-order AO is also assumed. Results are shown for satellites in 400-and 800km circular orbits. Figure 5 shows SKG rates calculated for a sky radiance of 25 W∕ðm 2 sr μmÞ and a 1 × HV 5∕7 turbulence profile plotted as a function of the receiver FOV. SKG rates are calculated for pointing angles of 0 deg, 30 Figure 5(a) shows results for the 400-km orbit with tilt correction. For FOVs larger than 15 μrad, the QBER due to sky noise is sufficiently high to preclude SKG. For FOVs below 15 μrad, sky noise can be reduced sufficiently to allow SKG within 45 deg of zenith. For a given FOV, however, key rates are unstable due to statistical variations in the uncompensated higher-order aberrations. With tilt correction alone, FOVs in the vicinity of 3 to 8 μrad represent the optimum trade-off between noise reduction and signal preservation.
With the addition of higher-order AO, shown in Fig. 5(c), stable SKG rates in excess of 1 kHz are possible within 60 deg of zenith. With higher-order AO, the standard deviation in SKG rates can be small relative to the mean value. With higher-order correction, FOVs in the vicinity of 2 to 3 μrad represent the optimum trade-off between noise reduction and signal preservation. Below this range, signal losses reduce key rates even with AO. At a FOV of 2 μrad, the Airy disk would theoretically pass the FS with Fig. 4 Examples of SKG rate PDFs calculated in the presence of turbulence assuming a receiver with a 2-μrad FOV pointing to zenith, a sky radiance of 25 W∕ðm 2 sr μmÞ, a turbulence profile of 1 × HV 5∕7 , a system rate of 10 MHz, and a satellite in a circular orbit at an altitude of (a) 400 km and (b) 800 km. Results obtained with tilt correction alone are shown in red. Results obtained with the addition of higher-order AO are shown in blue. about 84% power efficiency. Since AO is not perfect, the transmission is lower. Furthermore, the focused irradiance profile at the FS moves as AO iterates, leading to some time-dependent loss variance. Figure 5(b) shows results for the 800-km orbit with tilt correction. For a given elevation angle, the QBER as defined in Eq. (7) has increased due to the increased propagation losses. Consequently, SKG is only possible through further reductions in the FOV and at smaller angles relative to zenith where propagation distances are minimized for a given orbit altitude. Secure key rates in this case are subject to dropouts represented by large error bars relative to the mean value.
With the addition of higher-order AO, shown in Fig. 5(d), stable SKG rates in excess of 300 Hz are possible within 60 deg of zenith. As in the case of Fig. 5(c), the standard deviation in SKG rates can be small relative to the mean value.
Within each family of curves, SKG rates decline as the elevation angle increases from zenith. This is due to the increased losses that occur with beam divergence over the increased propagation distances and also due to the increased strength of turbulence that occurs with increased atmospheric path length. At 75 deg, the onset of deep turbulence and a high Greenwood frequency result in reduced AO system performance. Figure 6 shows the corresponding results calculated under more challenging conditions; namely, a sky radiance of 100 W∕ðm 2 sr μmÞ and a 3 × HV 5∕7 turbulence profile. Results for the case with tilt correction alone are shown in Figs. 6(a) and 6(b). With tilt correction alone, the signal loss due to turbulence at small FOVs is sufficiently high to preclude SKG at all elevation angles considered. SKG rates calculated with the addition of higher-order AO are shown in Figs. 6(c) and 6(d). With higher-order AO, SKG is possible. However, relative to the results shown in Fig. 5(c) and 5(d), the increased sky noise reduces the FOV at which SKG is possible.
For the case of the 400-km orbit shown in Fig. 6(c), SKG rates in excess of 1 kHz are possible at elevation angles within 45 deg of zenith for FOVs below about 7 μrad. Relative to the results in Fig. 5(c), the factor-of-four increase in H b requires a factor-of-two decrease in the FOV. Beyond 60 deg elevation angle, the effects of turbulence are stressing the capability of the AO system assumed in this particular simulation. At 60 deg from zenith, the Greenwood frequency, shown in Table 1, is 780 Hz, which is significantly larger than the 500 Hz AO correction bandwidth. At 75 deg from zenith, the Rytov value is 1.19 indicating deep turbulence where scintillation degrades SHWFS performance. At 75 deg, r 0 is only 65% of the SHWFS subaperture size, leading to poorly resolved wavefront tilts. The result is a reduced SKG rate at 60 deg and negligible SKG at 75 deg.
For the case of the 800-km orbit shown in Fig. 6(d), SKG is only possible at FOVs below 4 μrad. At this altitude, slew rates are lower and Greenwood frequencies are more benign than for the 400-km altitude case. However, the increased propagation distance results in increased losses that negatively impact key rates. SKG rates in excess of 200 Hz are possible within 45 deg of zenith, but rates are negligible beyond 60 deg from zenith.
Discussion
AO was demonstrated to be an enabling technology for daytime satellite-to-Earth QKD. The goal in implementing AO in a QKD receiver is to optimize SKG rates by optimizing the noise/loss trade space associated with the receiver FOV. Within a range bounded by turbulence and diffraction, AO is a mature technology for preserving the quantum signal at reduced FOVs. For example, assuming 1×HV 5∕7 turbulence, a 75 deg angle from zenith, and a perfect AO system, sky noise could be reduced by a factor of 400 without additional signal loss. AO, however, is not perfect. The performance is dependent upon the AO system parameters and the strength of turbulence. Photon capture efficiencies resulting from atmospheric turbulence and a closed-loop AO system were quantified for a specific system with numerical simulations based on Fresnel propagation and AO control theory. Information-theoretic estimates of SKG rates for a decoy-state QKD protocol were calculated based on the simulated capture efficiencies.
Results show that in moderate turbulence, simply reducing the receiver FOV to values smaller than previously discussed in the literature can reduce sky noise sufficiently to enable SKG in daylight. The addition of higher-order AO technologies enhances SKG rates considerably and even enables SKG in stronger turbulence where it would otherwise be prohibited as a consequence of background optical noise and signal loss due to turbulence and propagation. Furthermore, higher-order AO improves the stability of SKG rates when turbulence-induced losses are a factor. The relative benefit of AO is more significant at smaller FOVs where losses due to turbulence are greater.
The simulation can be applied to modeling the benefits of adaptive spatial filtering with other AO components and other turbulence conditions. Adaptive spatial filtering would likely benefit other QKD protocols and applications involving the transmission of quantum information over free-space channels. Adaptive spatial filtering may be of particular significance since time-bandwidth product considerations associated with Fourier transform and quantum uncertainty relationships limit the extent to which photons can be filtered spectrally and temporally. Fig. 6 SKG rates as a function of the receiver FOV calculated assuming a sky radiance of 100 W∕ðm 2 sr μmÞ, a turbulence profile of 3 × HV 5∕7 , and a system rate of 10 MHz. Numerical results are shown for pointing angles of 0 deg, 30 deg, 45 deg, 60 deg, and 75 deg from zenith assuming (a) a 400-km circular orbit with tilt correction, (b) an 800-km circular orbit with tilt correction, (c) a 400-km circular orbit with higher-order AO, and (d) an 800-km circular orbit with higher-order AO. | 10,308.2 | 2016-02-01T00:00:00.000 | [
"Physics"
] |
Modeling and Optimization in Investigating Thermally Sprayed Ni-Based Self-Fluxing Alloy Coatings: A Review
In investigating thermally sprayed Ni-based self-fluxing alloy coatings, widely applied under conditions of wear, corrosion, and high temperatures, designed experiments and statistical methods as a basis for modeling and optimization have become an important tool in making valid and comparable conclusions. Therefore, this paper gives an overview of investigating Ni-based self-fluxing alloy coatings deposited by thermal spraying by the use of designed experiments and statistical methods. The investigation includes the period of the last two decades and covers the treatments of flame spraying, high-velocity oxy/air fuel spraying, plasma spraying, plasma-transferred arc welding, and laser cladding. The main aim was to separate input variables, as well as measured responses, and to point out the importance of correct application of statistical design of experiment. After the review of the papers, it was concluded that investigators have used the process knowledge to analyze and interpret the results of the statistical analysis of experimental data, which is in fact the best way of using the design of experiment in every research. Nevertheless, more attention should be given to correct planning and conducting the experiments to derive the models suitable for the prediction of measured response and which could be an appropriate input for single- or multi-objective optimization.
Introduction
Ni-based self-fluxing alloys were developed in 1950s as alloys of nickel with added Cr, B, Si and C (hence, the often-used terms NiCrBSi, NiCrSiB, NiCrSiBFe, Ni-Cr-B-Si, etc.). They are applied in conditions of wear, corrosion and high temperatures. Boron and silicon (1 to 5 wt.%) added to nickel improve the fluxing properties and act as deoxidizers, forming borosilicate protecting other elements from oxidation [1,2]. In addition, boron and silicon (so-called temperature suppressants [2]), as well as chromium (10 to 20 wt.%), lower the melting temperature of pure nickel [2]. With the addition of chromium, good corrosion resistance is achieved, and thanks to the addition of carbon (up to 1wt.%), carbides are created, which, along with borides, silicides, carboborides, and some other phases in the nickel matrix, increase the hardness (even up to 70 HRC, depending on the chemical composition of the coating material). Iron, molybdenum, tungsten, and different hard particles or rare earth elements can be added, as well. With application in almost all areas of human activity (power plants, engines, conditions of extreme wear and corrosion, aerospace industry, molds and tools, machine parts, chemical industry, automotive industry, food processing industry, etc.) [1][2][3], these alloys can be an alternative
•
Recognition of and statement of the problem (being a teamwork it includes people well acquainted with the investigated process, statisticians, operators, laboratory technicians, etc.).
•
Selection of the response variable (what will be measured).
•
Choice of input variables (factors), levels, and ranges (what will be varied).
•
Choice of experimental design (as types of experimental designs are numerous, which one will be chosen depends on what the experiment is aimed at and which processing phase it is in).
•
Performing the experiment (three important principles, described below, are to be taken into the consideration). • Statistical analysis of the data (it is necessary to apply the knowledge of statistics).
•
Conclusions and recommendations ("from statistics" turn again to the investigated problem and use the process knowledge to explain statistical conclusions).
In addition to these seven steps, it is also necessary to apply three basic principles in conducting an experiment, originated from the above-mentioned Fisher. These are randomization, replication, and blocking (more detailed in Reference [17,18]). However, it happens sometimes that investigators do not apply these principles in experimental investigations (or do not know that they have to apply them) and, accordingly, this can lead to wrong conclusions. This has also been noticed by the review of literature on the application of designed experiments and statistical methods in investigating thermally sprayed Ni-based self-fluxing alloy coatings.
Therefore, in the present paper, the authors made a review of investigations of widely applied thermally sprayed Ni-based self-fluxing alloy coatings, including the statistical approach, in order to systematize input variables (what has been varied) and responses (what has been measured). The aim is also to point out the importance of observance of the three previously mentioned principles (randomization, replication, blocking) in experimental design and generally the importance of the correct planning and conducting of experiments and statistical analysis. The structure of most thermally deposited coatings can be very unpredictable since it is obtained by high-velocity deposition and cooling processes. It can be even dependent on the dimensions of the sample [7]. Consequently, in investigating these coatings, it is especially good to have at least the replication principle fulfilled. Furthermore, the present paper is a continuation of or addition to the research presented in authors' two-part paper [3,16] as regards thermally sprayed Ni-based self-fluxing alloy coatings. The collected important information on Ni-based self-fluxing alloy coatings was presented in these papers in a systematic way, and the diversity of research, for the period from 2000 to 2013, was pointed out, by citing 360 references. In connection with DOE methodology, this paper can be an addition to the research described in Reference [19,20]. In investigation Reference [19], Pierlot et al. reviewed the design and analysis of experiments methodology application in thermal spraying by comparing different experimental designs and evaluating their application. The author of Reference [20] also showed the most important types of the applied designed experiments in the field of thermal spraying, from screening experiments to those applied in the optimization phase. He also considered the methods of artificial intelligence (neural networks and fuzzy logic), showing their advantages and disadvantages with regard to the methodology of designed experiments. Through the review of papers, the author also proved the combination of designed experiments and methods of artificial intelligence where the designed experiments were in fact used as a basis to obtain the data that were afterwards processed by the methods of artificial intelligence.
After the previous general insight into the application of Ni-based self-fluxing alloy coatings and statistical design and analysis of experiments, the rest of the article is organized as follows: • In Section 2, the investigations in which the designed experiments are applied for obtaining regression models, reviewed, and presented. • Section 3 presents the research in which investigators did not apply (or did not specify) designed experiments for obtaining regression models or used some statistical methods.
•
The investigations in which the DOE methodology and methods of artificial intelligence are combined, or only DOE methodology is used with the main aim of optimization, are also covered in Section 4.
•
In Section 5, the discussion on the certain models, methodology, or optimization is presented. • Finally, the conclusions are provided in Section 6.
Statistical Modeling with DOE in Investigating Ni-Based Self-Fluxing Alloy Coatings Deposition
DOE methodology has important applications in the investigation of processes and products with the main aim of improving or optimizing their performance. In addition, it can serve as an important tool even in the development phase of a new process or product. Most authors have used the results of designed experiments for modeling, i.e., obtaining statistical (empirical, regression, mathematical) models for the prediction of response. For that purpose, researchers have frequently applied full factorial designs (factorials) where all levels of the factors are combined. Partial designs-fractions were used, as well, and some authors applied central composite design (CCD). Section 2.1 to Section 2.4 present the review of the papers where statistical modeling approaches applying DOE were used in investigating Ni-based self-fluxing alloy coatings deposition technologies.
PTAW and Laser Cladding
Dependence of response (mass loss) of PTA welded Ni-based self-fluxing alloys coatings (thickness 4-5 mm) at dry sliding wear-pin on disc method, on coating hardness (which depends on subsequent heat treatment), testing temperature, and sliding distance was obtained in Reference [21], by the application of full factorial design with three factors at three levels. It was concluded that testing temperature was a significant factor.
In the paper of Ramachandran et al. [22], authors applied central composite design to study wear rate at dry sliding wear of Ni-based self-fluxing alloys (along with cobalt-base alloy and stainless steel) deposited by PTAW on carbon steel (in thickness of 3 mm). The yield and ultimate tensile strength of Ni-based self-fluxing alloy was lower (310 and 655 MPa, respectively) than that of cobalt-base and higher than that of stainless steel. This does not apply to the microhardness that was highest for Ni-based self-fluxing alloy (480 HV0.1). Three factors were being changed (hardness and revolution speed of rollers and normal load at pin on roller test), each one at five levels. By statistical analysis of data, regression models were obtained with high coefficient of determination for the coating wear rate and the three previously mentioned coatings. It was concluded that the rollers' hardness was of considerable effect on response. Results and conclusions of the research, but only for Ni-based self-fluxing alloys, given in Reference [23], are similar to those in Reference [22]. Research of abrasive slurry of different PTA welded deposits of 3-mm thickness (Ni-based self-fluxing alloys, cobalt-base alloy, and stainless steel) was conducted in the paper by Ramachandran et al. [24]. Four different input variables were being varied at five levels (size of river sand particles in water, temperature and concentration of slurry, and number of revolutions of the samples), forming a central composite design of experiments with the aim to obtain statistical models of the wear rate dependence on the mentioned input variables. By increasing the size of river sand particles and the concentration of slurry (the highest influence), the wear rate was also increased, while, by increasing revolution speed of the samples, the wear rate was decreased. The slurry temperature variable had no considerable effect on the wear rate. Results and conclusions of research, but only for Ni-based self-fluxing alloys, given in Reference [25], are similar to those in the previously reviewed paper [24]. Ni-based self-fluxing alloy behaved different in sliding [22,23] and abrasive slurry [24,25] wear compared to cobalt-and iron-based alloy. This type of alloy had the lowest wear rate for sliding wear and the highest for abrasive slurry wear.
Reciprocating pin on plate sliding testing of 660-µm thick laser cladded and laser textured Ni-based self-fluxing alloy coatings was conducted by Garrido et al. [26]. They applied two sets of designed experiments. In the first experiment, at three levels, there were changes: the ratio between the surface covered by dimples and the whole texturing density and the geometric factor-the ratio between the diameter and the dimples depth (aspect ratio) aimed at defining the Stribeck curves. In the second experiment, instead of the first, previously mentioned factor, the distance between microdimples was defined at three levels, as well as combined with the microdimples diameter (as it was proved that the depth depends on the diameter) at three levels. The effect on the coefficient of friction change was observed with the aim to define the minimum one. It is demonstrated in this paper that experiments are an iterative process [17] in which the previously performed experiments are used to draw conclusions and design the next experiments.
Influence of three factors (laser power, powder feed rate, and scanning speed), changed at three levels, on the dimensions of Ni-based self-fluxing alloy coating (clad height, clad width, and clad penetration depth into substrate) was researched in paper by Davim et al. [27] with the aim to obtain a regression model. For the confirmation experiment, the authors compared actual and predicted values and calculated the error, which was acceptable for clad height and width (7.6% and 6%, respectively), while, for the penetration depth, it was high (20.1%).
For the needs of numerical modeling (by finite element method) and aimed at estimating and comparing the height and penetration of Ni-based self-fluxing alloy coating and heat affected zone, Felde et al. [28] conducted a full factorial design changing two factors (scanning speed and laser power) at four levels. They concluded that, in some cases, there is a difference between the coating estimated dimensions obtained by simulation (numerical modeling) and the dimensions of real samples because all possible parameters of influence on the complicated process of laser alloying for modeling were not considered.
Garrido et al. [29] studied laser texturing of laser cladded Ni-based self-fluxing alloy coatings (>660 µm thick, of 800 to 880 HV0.3 microhardness), by applying central composite design. They changed three input variables (energy, laser spot diameter, and impulse duration) with the main aim to obtain statistical dependence of the diameter and the dimples' depth on the previously mentioned input parameters.
Laser single-pass and multi-pass cladding of NiCrBSi+WC coatings was investigated by Dubourg and St-Georges [30], with the aim of obtaining statistical dependence between laser cladding parameters (scanning speed, powder feed rate, overlapping, and out-of-focus distance) and width, height, and penetration depth of the laser clad, as well as the content of WC in the coating. The Taguchi experimental design was combined with the experimentation and modelization design of experiments. The second mentioned method is an interactive technique that enables a simultaneous varying of all factors. Scanning speed and powder feed rate were significant factors influencing the clad geometry.
Flame Spraying
Resistance to abrasive wear of flame sprayed and simultaneously fused Ni-based self-fluxing alloy coatings (NiCrBSi and NiBSi) of 1-mm thickness, with the application of dry sand/rubber wheel test, was investigated by Havrlisan et al. [10]. By applying full factorial design 2 2 with three replicates, Materials 2020, 13, 4584 6 of 27 for two standardized variants (100 and 2000 revolutions of wheel), authors obtained the regression models, which showed dependence of volume loss on spraying distance and type of applied coating. For the shorter variant of testing, it was proven that the spraying distance, along with the type of coating, was significant, as well the interaction between these two factors (Figure 1a), while, for the longer variant, only the type of coating was significant (Figure 1b). Authors concluded that this might be connected with the microstructure in the coating surface and greater data variability for repetitions in longer variant and affected different coating layers with different structures. In that way, the authors of this paper (who are also the authors of the present investigation) explained and corroborated the statistical conclusions by process knowledge and thus complied with the last step in the guidelines for designing the experiment [17], shown in Section 1.
Materials 2020, 13, x FOR PEER REVIEW 6 of 27 in longer variant and affected different coating layers with different structures. In that way, the authors of this paper (who are also the authors of the present investigation) explained and corroborated the statistical conclusions by process knowledge and thus complied with the last step in the guidelines for designing the experiment [17], shown in Section 1. In their research of flame sprayed and furnace fused Ni-based self-fluxing alloy coatings, Bergant et al. [31] applied mixed-level factorial design varying the factor time of fusing at two levels and the factor temperature of fusing at three levels. The conclusion they reached was that the lowest porosity was at temperature 1080 °C and time of 10 min. While analyzing variance, it was concluded that the porosity was influenced by the temperature and time of subsequent heat treatment, as well as that interaction between factors was low. After the fusion process at 1080 °C for 20 min, the coating thickness decreased (from an average of 445 µm for a sprayed state to an average of 329 µm), due to the increase in density and entrapped gases release; in addition, the surface roughness Ra decreased In their research of flame sprayed and furnace fused Ni-based self-fluxing alloy coatings, Bergant et al. [31] applied mixed-level factorial design varying the factor time of fusing at two levels and the factor temperature of fusing at three levels. The conclusion they reached was that the lowest porosity was at temperature 1080 • C and time of 10 min. While analyzing variance, it was concluded that the porosity was influenced by the temperature and time of subsequent heat treatment, as well as that interaction between factors was low. After the fusion process at 1080 • C for 20 min, the coating thickness decreased (from an average of 445 µm for a sprayed state to an average of 329 µm), due to the increase in density and entrapped gases release; in addition, the surface roughness Ra decreased from an average of 9.88 µm to an average of 3.26 µm). The lowest average porosity of 0.87% was obtained after furnace fusing at 1080 • C for 10 min.
Resistance to abrasive wear (pin on disc test) of flame sprayed Ni-based self-fluxing alloy coatings (of 0.95 to 1.05 mm thickness) with addition of 0.4% of CeO 2 and 0.6% of La 2 O 3 was analyzed while changing four factors (size of abrasive, loading, temperature, sliding speed) at three levels [32]. This author also calculated the difference between actual (measured) and predicted (obtained by the statistical models) values, i.e., error. In contrast to the usual amount of chromium (about 15%), a Ni-based self-fluxing alloy with 1.2 to 1.6% chromium was used here. The porosity decreased from 6.2 to 5.4% due to the addition of rare earth oxides. In addition, rare earth oxides influenced the increase of hardness (from 210 ± 9 HV5 to 241 ± 12 HV5) due to the refinement of structure and the higher amount of eutectic.
Plasma Spraying
Fernandez et al. [33] investigated the influence of four input variables at two levels-load, amount of hard WC particles in the coating, size of abrasive testing particles, and type of testing (dry and wet)-on the abrasion mass loss. The thickness of the coatings was 2-3 mm, and hardness of 855 ± 20 HV0.3 was obtained for coatings without WC, while, for the coatings with WC added, the hardness of the matrix was 850 ± 20 HV0.3, and, for the WC, it amounts to 1290 ± 20 HV0.3. It was proven that the size of abrasive testing particles and presence of hard particles in the plasma sprayed and remelted Ni-based self-fluxing alloy coating are of great influence to abrasive wear.
Valente [34] applied factorial design in investigation of controlled atmospheric plasma sprayed and induction remelted coatings of 300 µm thickness by varying the process parameters, power and spraying distance, at two levels, and the third factor, powder volumetric fraction (NiCrBSi and Mo powder), at three levels. The author concluded that the coating containing 10% Mo and 90% NiCrBSi powder had low porosity (0.9%), high hardness (668 HV0.3), good corrosion resistance, and good properties of wear.
Load, testing temperature, presence of WC in coating, and deposition technologies were varied at two levels in the investigation of Rodriguez et al. [35] when they applied reciprocating pin on plate sliding wear test for Ni-based self-fluxing alloy coatings of 300 to 500 µm thickness, plasma, and flame sprayed with remelting. The hardness of plasma sprayed coatings was 50 to 55 HRC, while for the flame sprayed and fused it was 60 HRC. After fusing, the surface roughness decreased to 2-4 µm, when comparing to flame sprayed state (20-30 µm). All factor levels combinations were tested, i.e., the full factorial design was applied with responses-the mass loss of coatings but also of counter body (alumina) after sliding wear testing. After statistical analysis of the wear data, it was concluded that the significant factors were the kind of applied thermal spraying and load. That the two other factors, temperature and WC particles, were not significant, was explained by the fact that Ni-based alloys were wear resistant at higher temperatures and that due to the bad connection between WC and the matrix the WC particles were pulled out of the matrix; thus, they did not contribute to the higher resistance to wear.
Tribological properties of plasma and HVOF sprayed Ni-based self-fluxing alloy coatings with solid lubricant Fe 2 O 3 were studied by Zorawski and Skrzypek [36]. For plasma spraying, the input variables were the amount of Fe 2 O 3 , spraying distance, gas pressure, and current, while, for HVOF treatment, along with the amount of Fe 2 O 3 and spraying distance, the flow rate of oxygen and propane was also changed. The responses were as follows: surface roughness, coefficient of friction, and microhardness. For the mentioned properties, for both spraying treatments, by statistical analysis of experimental data obtained by the fractional design of experiment, the authors developed linear regression models. Spraying distance and amount of Fe 2 O 3 solid lubricant showed as significant factors of influence on Materials 2020, 13, 4584 8 of 27 surface roughness, but for the HVOF procedure only, while the amount of Fe 2 O 3 solid lubricant was a significant factor for the coefficient of friction of the HVOF deposited coatings. For plasma sprayed coatings, the maximum microhardness was 642 HV0.5, and, for HVOF, sprayed coatings, it was 706 HV0.5. Regarding the surface roughness Ra, the minimum value was 0.42 µm for plasma spraying, and, for the HVOF procedure, the minimum surface roughness was 0.18 µm. A lower porosity was achieved for the HVOF process (1.13 ± 0.56%) than with the plasma spraying process (3.19 ± 1.76%).
HVOF Spraying
Resistance to pin on disc wear and corrosion of three types of HVOF sprayed coatings (NiCrBSi, WC-12%Co, and Cr 3 C 2 -25%NiCr) of 150 to 170 µm thickness was investigated by Shabana et al. [37]. For the pin on disc wear examination they performed mixed-level factorial design changing two factors (load and temperature) at two levels, while the third factor (sliding distance) was changed at five levels. Statistical models of the wear dependence on the mentioned factors were obtained with high coefficients of determination. When comparing these three types of coatings, NiCrBSi coatings had the lowest adhesive strength (58.15 MPa) and lowest microhardness (997 HV0.05), while the porosity was lower (amounts to 0.92%) than the porosity for Cr 3 C 2 -25%NiCr (1.49%).
Gisario et al. [38] examined laser remelting of HVOF deposited Ni-based self-fluxing alloy on aluminium alloy substrate. Mixed-level full factorial experiment was conducted changing two factors, laser power and rotational scan speed at six and five levels, respectively, with two replications of factor levels combinations. Effectiveness index of laser remelting was the response for which the nonlinear regression model was obtained, providing the conclusion that a better ratio will be achieved by higher laser power and smaller rotational speed. Effectiveness index was defined as the ratio of the thickness of the coating affected by laser remelting and the total thickness of the coating.
Analysis of resistance to abrasive wear (pin on disc test) of HVOF Ni-based self-fluxing alloy coatings with addition of WC and CeO 2 was conducted in Reference [39]. Due to the addition of 0.4% CeO 2 , the microhardness was increased from 1053 ± 113 to 1185 ± 96 HV0.1. Pointing out that a number of authors use experimental "one-factor-at-a-time" strategy [17], the author changed four factors at three levels (size of abrasive, load, temperature, sliding distance), with the aim to obtain important factors to response, average mass loss, but also a statistical model of abrasive wear that shows the dependence of response on input variables. Verification experiments were also conducted so as to verify the model in which loading, size of abrasive, and sliding distance were significant factors, with the first order interaction between load and size of abrasive, load, and sliding distance and between size of abrasive and sliding distance.
In Table 1, the investigations described in Section 2.1 to Section 2.4 are listed according to deposition technology and applied experimental designs. Size of abrasive; Load; Temperature; Sliding distance Mass loss (pin on disc two-body abrasive wear) Fractional factorial design [39] PTAW-plasma-transferred arc welding; LC-laser cladding; FS-flame spraying; PS-plasma spraying; HVOF-high-velocity oxy fuel spraying.
Statistical Modeling and Different Statistical Methods in Investigating Ni-Based Self-Fluxing Alloy Coatings Deposition
There are authors who did not apply designed experiment (or did not mention this fact or there are no experimental data to make conclusion about the type of experimental design). They used statistical analysis of data from unplanned or unspecified experiments (or measurements) to obtain regression models (sometimes curve fitting term was used) or used some statistical methods to process and analyze the data without obtaining a regression model (e.g., Weibull distribution, descriptive and interferential statistics parameters).
The papers with statistical modeling and application of different statistical methods in investigating Ni-based self-fluxing alloy coatings deposition technologies are described in Section 3.1 to Section 3.4.
PTAW and Laser Cladding
Flores et al. [40] investigated erosion-corrosion resistance of PTA welded Ni-based self-fluxing alloy coatings with added WC varying the sand content in slurry, as well as temperature and velocity of slurry. Statistical dependences of the total mass loss (TML) on the slurry velocity (v) was obtained with high coefficients of determination. For Ni-based self-fluxing alloy coating, the statistical dependence, for the temperature 65 • C, was as follows: TLM = 0.0036 . v 3.1222 with the coefficient of determination of 0.9984.
Badisch et al. [41] studied the influence of microstructure (average distance between WC particles, average diameter of WC particles and volume fraction of WC particles, matrix hardness, and shape of hard particles) of composite Ni-based self-fluxing alloy coatings obtained by PTAW and laser cladding on continuous impact abrasion wear. It was shown that the matrix hardness, along with the average diameter of WC particles, average distance between WC particles, and volume fraction of WC particles, had a great influence on continuous impact abrasion wear rate, while the shape of hard particles in the coating was of no significance.
Block on ring tests of resistance to sliding wear were applied in Reference [42] on laser cladded NiCrBSi coatings deposited to grey cast iron substrate. These authors obtained a statistical linear model for evaluation of the wear mass loss dependence on volume removed (coefficient of determination was 0.9962). By the statistical linear model Fernandez et al. [42] also evaluated average wear rate in dependence on the product of normal load and sliding speed (coefficient of determination was 0.8638). These authors measured the microhardness by coating depth at three different locations. The average microhardness was 900 HV0.3, with a decrease in hardness at the substrate/coating interface, due to, as the authors claim, the diffusion of Fe from the substrate. The decrease in hardness was more pronounced at the edges of the sample.
In studying the resistance to sliding wear of laser cladded NiCrBSi+WC coatings of 500 HV1 matrix hardness, Garcia et al. [43] obtained the exponential statistical dependence (with R 2 = 0.8503) of the wear track cross section on the fraction of WC hard particles for different speeds of sliding, showing that the increase of the quantity of hard particles in a coating results in the decrease of wear. However, these authors mention marginal fraction of 27% of hard particles (actual concentration), above which there was not any significant reduction of wear. The authors measured the actual concentration of WC particles in the coating relative to the concentration in the feeder.
Linear single and multiple regression models (with high coefficients of determination) of abrasive wear rate dependence on parameters connected with WC particles (average distance between WC particles, average diameter of WC particles, and volume fraction of WC particles) in dry sand/rubber wheel abrasive test were developed in a study on laser cladded Ni-based self-fluxing alloy coatings, by Polak et al. [44]. They proved that matrix hardness did not affect mass loss while average distance between WC particles was of greatest significance (different from the study in Reference [41], where the matrix hardness was a significant factor for continuous impact abrasion wear).
Coefficient of determination for statistical linear dependence of hardness of a Ni-based self-fluxing alloy coating cladded by laser on stainless steel on the amount of added hard WC particles was presented in Reference [45]. The microhardness of the matrix increased from 300 HV0.3 to 350 HV0.3 when a smaller amount (5%) of WC particles was added, and, if a larger amount (45%) of WC particle was added, then the microhardness of the matrix increased to 900 HV0.3. These authors have the impact of wolfram carbide particles on intensity-laser-induced breakdown spectroscopy (method for characterization of coatings) presented by linear regression model, with the coefficient of determination of 0.999.
In Reference [46], which presents the results of investigation of NiCrBSi coatings deposited by cladding with CW CO 2 laser, for three different particle sizes of NiCrBSi powder (smaller than 20 µm, from 20 to 80 µm and from 80 to 100 µm), the response surfaces were obtained that show the dependence of powder feed rate (g/min) on the pressure of carrier gas-air (MPa) and on the gas flow (dm 3 /min). They concluded that the pressure and the air flow increase resulted in the increase of the powder feed rate, but not equally for all particle sizes of the powder. Therefore, to continue the research, they used the powder particles of size 20 to 80 µm while the input variables were the cladding distance (changed at three levels, 10, 12, and 14 mm) and the laser spot speed (changed at five levels, 40, 60, 80, 100, 120 mm/min), while the response were the geometrical features of single laser track-height and width. Based on the conducted experiments and obtained regression models it was concluded that the increase of cladding distance resulted in the decrease of width and the increase of height, while the increase of laser spot speed resulted in the decrease of width and height of clad. The authors [46] think that the temperature gradients are the key factors for this.
Sliding wear (pin on disc) and Knoop hardness of laser cladded NiCrBSi coatings exhibited to aging in salt fog was investigated and the influence of aging time on mechanical properties was presented in Reference [47]. The authors used Weibull statistical analysis to analyze the Knoop hardness data, and, consequently, Weibull modulus versus aging time was obtained. For the aging time up to 800 h, the structure of the coating was homogenous (without pitting or crevice corrosion), and the wear properties were excellent (higher Weibull moduli were obtained).
Applying graphical software and statistical analysis (descriptive statistics) in the investigation, Ma et al. [48] proved that after laser remelting of previously laser cladded NiCrBSi coating with 50% added WC particles, the median size of particles was reduced from 35.4 to 5.62 µm. After laser remelting, there was a decrease in microhardness by 50 HV0.1 compared to microhardness after laser cladding (which was 900 to 1050 HV0.1).
Flame Spraying
Similar to Reference [42], a block on ring tests of resistance to sliding wear was applied in Reference [49], but, for flame spraying and flame remelting (R 2 was 0.9822), as well as for laser remelting (R 2 was 0.9874) for the dependence of mass loss on volume removed. After flame and laser remelting, the microhardness of the coating was approximately the same and was 900 HV0.1, which was lower than for flame spraying condition (1100 HV0.1).
In the paper of Stanford and Jain [50], the dependence of wear volume on wear time at pin on disc sliding test for flame sprayed NiCrBSi alloy coating (final thickness of 0.26 mm was achieved after grinding) is shown by a linear regression model with a high coefficient of determination R 2 (0.9968), while the dependence of coefficient of friction on time could not be presented by a regression model with high coefficient of determination due to dissipation of data. The amount of porosity was 2.1%, and the hardness was 789 HV.
Sliding test with a lubricant into which debris particles (alumina Al 2 O 3 ) were added was conducted for investigating flame sprayed and flame remelted NiCrBSi coatings of 1.1 mm thickness, in Reference [51]. Instead of initial surface roughness, authors suggested applying actual surface roughness in expressions for obtaining the coefficient of friction (Stribeck curves). For investigation, they used three different sizes of Al 2 O 3 particles (1.5, 32.5, and 45 µm) and five kinds of normal load of 3, 6, 9, 15, and 20 N, and, for each load, the sliding speed was also changed. Curve fitting was applied for obtaining statistical dependence and coefficient of determination for Stribeck curves (coefficient of friction dependence on so-called lubrication number parameter). Included into this parameter was the initial surface roughness, lubricant viscosity, sliding speed, and contact pressure. With the aim to include the actual surface roughness, they also obtained the fitted curves with excellent coefficient of determination (0.9902) for the surface roughness dependence on debris particle size, and they included this expression, instead of initial roughness, into the expression for obtaining the coefficient of friction's dependence on such new lubrication number into which the actual surface roughness was included.
Plasma Spraying
Fernandes et al. [52] statistically processed data to obtain confidence intervals for specific wear rate of plasma sprayed Ni-based self-fluxing alloy coatings with addition of ceramic ZrO 2 particles; besides, using regression model, linear dependence was shown between frictional force and loading so as to compute the coefficient of friction, which was proven to decrease with ceramic particles added. These authors investigated coatings applied using two premixed powders and the dual system of feeding.
Atmospheric plasma spraying of NiCrBSi coatings with two sizes of powder particles (50 to 75 µm and 75 to 100 µm) was investigated in Reference [53]. By applying powder with smaller particles, a greater coating thickness (598 ± 24 µm) was achieved than the thickness (432 ± 20 µm) obtained by applying powder with larger particles. This was explained by the fact that larger particles sometimes cannot be melted and weaker adhering can be achieved. In investigating electrochemical corrosion in 3.5 wt.% NaCl solution, statistical approach was applied for obtaining statistical models of the following electrochemical impedance spectrum (EIS) data: solution, film, and charge transfer resistance, capacity element, Warburg impedance, and constant phase element. Chi-square test was used to compare the actual (measured), and the predicted (calculated) data and good correlation was proven.
Porosity of plasma sprayed Ni-based self-fluxing alloy coatings was the subject of research in Reference [54]. The aim was to obtain correlation and dependency between linear (by length) and the corresponding spatial (by volume) dimensional parameters of pores in a coating. Author Das [54] points to the need to evaluate 3D parameters of pores based on the analysis of the sample's cross section, where the pores are represented as circles, while, in reality, they are pores in which the pore diameter cannot be well represented by the diameter of the cross-section circle as it depends on the section plane. Analysis of linear and spatial parameters was performed on plasma samples but also on the plasma and fused (sintered, as the author states) samples. Based on the calculated spatial parameters of the pores defined by linear parameters obtained on the sample's cross section, it was concluded that the sintered plasma coatings had 100 to 400 times higher quantity of pores, but the relative porosity was lower due to the much smaller diameters of pores. After the sintering procedure, the relative spatial porosity decreased from 21-34% to 2.06-3.07%, and the pore diameter decreased from 350 µm to 38.4 µm.
From the statistical point of view, porosity of plasma sprayed Ni-based self-fluxing alloy coatings was analyzed in Reference [55]. Authors first checked whether the porosity data were normally distributed or according to the Weibull distribution. Additional statistical analysis was conducted in accordance with the Weibull distribution, and it was proven that the porosity of coatings was reduced as the hydrogen flow rate was increased. Authors' explanation for this was higher temperature and speed of molten particles, therefore better bonding between the molten particles themselves in the coating. They also mentioned that increase of the hydrogen flow rate resulted in reduction of the residual stresses among molten particles. The third reason they gave for reduced porosity was also the captured air in which quantity was reduced with increased hydrogen flow rate. Based on dispersion of the data on porosity for lower hydrogen flow rates, it was concluded that this was the result of inhomogeneity of structure (high pores unevenly distributed). Investigations presented in Reference [55] were continued with the application of the same statistical approach, and the results were presented in Reference [56]. The authors concluded that the porosity was increased at lower spraying power due to microcracks and unmelted particles. They also mentioned the limit power of 57 kW above which the quantity of cracks was not practically reduced any longer because all of the particles were melted already. Spraying power influence on the porosity but also mechanical properties (modulus of elasticity, microhardness, and residual stresses) was also investigated [57]. In Reference [58], the authors were varying powder feed rate and concluded that, at higher powder feed rates, there was quite a number of unmelted particles and inhomogeneity in the plasma sprayed coating leading to the increase of porosity, but the modulus of elasticity, microhardness, and residual stresses were reduced till some minimum value and then increased with increasing powder feed rate. In Reference [59], the authors gave the results of investigations of porosity, modulus of elasticity, and microhardness related to the plasma spraying parameters (power, powder feed rate, and hydrogen flow rate) by using the same statistical approach as in Reference [55][56][57][58].
Frequently applied statistical Weibull distribution function for analysis of life data was used in investigations of Zhang et al. [60] and Zhang et al. [61], who studied rolling contact fatigue of plasma sprayed and laser remelted NiCrBSi coatings. In Reference [60], authors conducted 13 experiments in order to obtain the Weibull plot that shows predicted failure probability for different fatigue life data.
HVOF Spraying
Microstructure, hardness, modulus of elasticity, yield strength, surface roughness, and surface residual stresses of HVOF sprayed Ni-based self-fluxing alloy coatings of 230 µm thickness were investigated depending on spraying distance, in Reference [62]. Authors concluded that the increase of spraying distance resulted in the increase of the number of unmelted particles by volume, increase of surface roughness (from 8 to 11 µm) and decrease of the modulus of elasticity, while the linear statistical model proved the dependence of surface roughness Rz on the volume fraction of unmelted particles. These authors also presented the statistical dependence of interplanar spacing on sin 2 ψ (XRD measurement), as well as the dependence of indentation load on penetration depth. The higher compressive residual stresses for smaller spraying distances were also proven.
For dry sand/rubber wheel abrasion test, Miranda and Ramalho [63] obtained linear regression models of abrasive wear dependence on volume fraction of hard particles (WC+Co) for HVOF and flame sprayed Ni-based self-fluxing alloy coatings (of 0.5 mm thickness) applicable for the 30 to 70% fraction of hard particles. HVOF coatings had lower porosity (0.5%) and higher adhesive strength than flame sprayed coatings. The only exception is the flame sprayed and remelted coating with 30% WC+Co, for which the porosity was also only 0.5%. Similar to Reference [48], Rukhande and Rathod [64] used descriptive statistics to obtain important parameters (mean, standard deviation, and median) for NiCrSiBFe powder particles and their distribution. They applied laser diffraction method for analysis of powder particles. The porosity was 2.21% for HVOF coatings and 3.09% for plasma sprayed coatings.
The results of investigating correlation between the properties of in-flight particles (temperature, velocity, and size) and the properties of plasma, flame, and HVOF sprayed NiCrBSi coatings (amount of oxides, porosity, adhesive strength, hardness, and modulus of elasticity) are presented in Reference [65]. The authors obtained the same coating thickness of 300 µm for all three procedures. The authors concluded that for the HVOF deposition technology the best properties achieved due to highest velocities of in-flight particles, i.e., the highest kinetic energy of particles at impact resulting in strong cohesion and hardness of the coating. The statistical processing of the collected data was performed to obtain the average values and standard deviations. The velocity of particles has been shown to depend on their diameters. The larger the diameter, the lower the speed. This is most pronounced in the HVOF process, where velocities of 500 and 400 m/s were achieved for smaller and larger particles. In the plasma spraying procedure, the velocities were 160 and 120 m/s. For the flame spraying process, the velocities were the lowest (41 and 35 m/s) and did not differ significantly for larger and smaller particle diameters.
Those type of investigations reviewed in Section 3.1 to Section 3.4 are summarized in Table 2, following the classifying principle from Section 2.
Optimization in Investigating Ni-Based Self-Fluxing Alloy Coatings Deposition
To optimize a process or product design, response surface methodology (RSM) [17][18][19][20]66] and Taguchi approach [19,20,67,68] have often been using. Along with the DOE methodology, the artificial intelligence methods are also applied in thermal spraying among other methods for optimization, as shown in a review paper [20], previously cited.
In the following Sections, the papers with optimization approaches in investigating Ni-based self-fluxing alloy coatings deposition technologies are presented.
PTAW and Laser Cladding
In the research by Siva et al. [69] dealing with plasma transferred arc welding (PTAW) of Ni-based self-fluxing alloy to stainless steel, authors varied the travel speed, current, oscillation amplitude, preheat temperature and powder feed rate at five levels, aimed at predicting weld bead dimensions and dilution and at obtaining optimal parameters, by applying the combination of designed experiment (central composite design, CCD) and genetic algorithms, GA. In Introduction, these authors point out the precedence of application of genetic algorithms over other optimization methods, particularly the method of steepest ascent (descent) that is applied in response surface methodology (RSM) when reaching the region of optimal parameters. Optimal values of PTAW parameters for obtaining minimum depth of penetration to substrate, maximum height and width of weld bead and minimum dilution, obtained by the application of GA approach (minimum penetration 0.317 mm; maximum height 4.165 mm; maximum width 23.973; and minimum dilution 5.632%), proved better than the values obtained by Generalized Reduced Gradient (GRG) method (minimum penetration 0.36 mm; maximum height 4.115 mm; maximum width 23.973; and minimum dilution 6.356%), built into the Excel Solver and applied on non-linear models.
Tu et al. [70] aimed at finding optimal parameters of PTA welding for two types of coatings (Stellite cobalt alloy and Ni-based self-fluxing alloy) on 0.48% carbon steel. Two methods were applied and compared, the Taguchi method and the Taguchi regression method for obtaining optimal values of accelerating voltage, powder feed rate, and preheat temperature (these factors were identified as significant), as well as of current, waving oscillation, plasma gas rate, and rotation, which will give minimal sliding (pin on disc) wear mass loss. For both applied methods, Stellite was in optimal combination and the Taguchi regression method gave better signal to noise ratio than the Taguchi method. When comparing the actual and predicted values of mass loss, i.e., signal to noise ratios, for the Taguchi method, an average error of 7.05% was obtained, and for the Taguchi regression method, the error was 5.55%, Three different responses (volume fraction of WC/W 2 C particles in the coating, hardness of the coating matrix, and equivalent diameter of WC/W 2 C particles) depending on three different input variables (parameters) of PTA welded Ni-based self-fluxing alloys were investigated by Ilo et al. [71] using the grey relational Taguchi method with the aim to obtain optimal parameters (welding current, welding speed and oscillation speed). The authors concluded that it is important to determine optimal parameters because of the considerable influence they have on the dissolution of carbides in Ni matrix. By reducing the volume fraction of WC/W 2 C particles in the matrix (because of dissolution), the matrix hardness will be increased, but, at the end, the resulting hardness will be decreased because of lower microhardness of new phases, which can be of influence to resistance to abrasive wear. Depending on the combination of process parameter levels, the minimum matrix microhardness was 518 HV0.1, while the maximum value was 785 HV0.1.
Laser cladding of Ni-based self-fluxing alloy on steel with 0.95% carbon was investigated by Onwubolu et al. [72] who applied the combination of factorial design and artificial intelligence (scatter search optimization as a method of evolutionary algorithms). The response was the clad angle (which must be large enough to avoid porosity, according to the authors), calculated by means of the clad width and height, input variables being scanning speed, laser power and powder feed rate, whose optimal values after application of scatter search algorithm were the following: scanning speed 5 mm/s, power 2 kW and powder feed rate 5 g/min.
Wu et al. [73] studied the porosity of laser cladded NiCrBSi coatings mainly with the aim of obtaining a statistical model and its application for optimization. Input variables were laser power, scanning speed and powder feed rate, each varied at three levels. Response was the porosity area that was set during optimization to be minimal. They have proven that the most significant factor was powder feed rate. With optimal parameters, laser power of 1524.80 W, scanning speed of 6.72 mm/s, and powder feed rate of 5.20 g/min, a porosity rate of 0.01% was achieved. Their study also proved that porosity could not be completely avoided by the application of the suggested optimal parameters, and concluded that the powder without porosity should also be used, as well as such a deposition process applied, which would not require a high shielding gas flow.
Flame Spraying
The application of Taguchi method for obtaining optimal values of substrate surface roughness and preheat temperature, oxygen-acetylene ratio (type of flame) and spraying distance is presented in References [74,75], aimed at obtaining maximum adhesive strength of flame sprayed and furnace remelted Ni-based self-fluxing alloy coatings on low carbon steel. Optimum values, verified by a new confirmation experiment, as well, were as follows: substrate material surface roughness 9.2 µm, substrate preheat temperature 200 • C, spraying distance 200 mm, and flame type-carburizing with obtained average adhesive strength of 22.35 MPa.
By the use of Taguchi method, cracking resistance of three types of flame sprayed and simultaneously fused NiCrBSi coatings was investigated and the results are presented in Reference [8]. The author proved that NiCrBSi+WC coating (P2) of thickness 0.8 mm (D3) flame sprayed and simultaneously fused on steel C45 (M1) should be most cracking resistant in three-point bending test (Figure 2). In order to check the optimal combination, the author conducted a confirmation experiment for which an average critical force of 33 kN was obtained. Wu et al. [73] studied the porosity of laser cladded NiCrBSi coatings mainly with the aim of obtaining a statistical model and its application for optimization. Input variables were laser power, scanning speed and powder feed rate, each varied at three levels. Response was the porosity area that was set during optimization to be minimal. They have proven that the most significant factor was powder feed rate. With optimal parameters, laser power of 1524.80 W, scanning speed of 6.72 mm/s, and powder feed rate of 5.20 g/min, a porosity rate of 0.01% was achieved. Their study also proved that porosity could not be completely avoided by the application of the suggested optimal parameters, and concluded that the powder without porosity should also be used, as well as such a deposition process applied, which would not require a high shielding gas flow.
Flame Spraying
The application of Taguchi method for obtaining optimal values of substrate surface roughness and preheat temperature, oxygen-acetylene ratio (type of flame) and spraying distance is presented in References [74,75], aimed at obtaining maximum adhesive strength of flame sprayed and furnace remelted Ni-based self-fluxing alloy coatings on low carbon steel. Optimum values, verified by a new confirmation experiment, as well, were as follows: substrate material surface roughness 9.2 μm, substrate preheat temperature 200 °C, spraying distance 200 mm, and flame type-carburizing with obtained average adhesive strength of 22.35 MPa.
By the use of Taguchi method, cracking resistance of three types of flame sprayed and simultaneously fused NiCrBSi coatings was investigated and the results are presented in Reference [8]. The author proved that NiCrBSi+WC coating (P2) of thickness 0.8 mm (D3) flame sprayed and simultaneously fused on steel C45 (M1) should be most cracking resistant in three-point bending test (Figure 2). In order to check the optimal combination, the author conducted a confirmation experiment for which an average critical force of 33 kN was obtained. Figure 2 is a common presentation after applying the Taguchi method. Taguchi defined the evaluation method, which is S/N ratio (signal-to-noise ratio). The combination of the levels of control factors that has the highest S/N ratio is optimal, i.e., the construction is more robust, i.e., less sensitive to noise factors. In investigation Reference [8], the expected S/N ratio was 31.077 ± 1.88, and the calculated S/N ratio after the confirmation experiment was 30.35 and was in the expected range of 29.19 to 32.95. Figure 2 is a common presentation after applying the Taguchi method. Taguchi defined the evaluation method, which is S/N ratio (signal-to-noise ratio). The combination of the levels of control factors that has the highest S/N ratio is optimal, i.e., the construction is more robust, i.e., less sensitive to noise factors.
In investigation Reference [8], the expected S/N ratio was 31.077 ± 1.88, and the calculated S/N ratio after the confirmation experiment was 30.35 and was in the expected range of 29.19 to 32.95.
The guidelines for designing the experiment listed in Section 1 of the present paper were followed in Reference [8], which is presented in Figure 3. The guidelines for designing the experiment listed in Section 1 of the present paper were followed in Reference [8], which is presented in Figure 3.
Plasma Spraying
Two factors, percentage of laser remelted surface and the angle of laser meshing were changed at four levels, in the investigation of Vijande et al. [76], who studied plasma sprayed and partially laser remelted Ni-based self-fluxing alloy coatings. The result of performed full factorial design was a linear regression model with a high coefficient of determination, which showed wear mass loss dependence on the two previously mentioned factors. Then, the input variables optimal values (50% of laser remelted surface and the angle of meshing of 22.5 • ) were calculated to obtain minimal mass loss at lubricated wear.
Liu et al. [77] optimized the parameters (powder feed rate, power, argon and hydrogen flow rate) for supersonic plasma spraying, applying the Box-Behnken designed experiment. Response was the NiCrBSi coating porosity. Optimal values for input variables were the following: 90 g/min (powder feed rate), 50 kW (power), 3.1 m 3 /h (flow rate of argon), and 0.35 m 3 /h (flow rate of hydrogen), and estimated porosity was 1.8%.
In addition to the powder feed rate, power, and plasma gas flow rate, Dubrovskaya et al. [78] optimized some more plasma spraying parameters (spraying distance, rotational speed of the specimen, and the displacement of the spraying torch) by the application of the steepest descent method in investigating Ni-based self-fluxing alloys with the addition of zirconium and hafnium. The responses were coating microhardness, porosity and thickness, the presence of cracks, and adhesive strength. The authors calculated the optimal values of the input variables, which were as follows: spraying distance 100 mm, rotational speed of the specimen 20 rpm, displacement of the spraying torch 10 mm/min, powder feed rate 40 g/min, power 36 kW, and plasma gas flow rate 50 l/min.
HVOF Spraying
Rukhande and Rathod [64] applied the Taguchi method to investigate the influence of the type of deposition technology (HVOF and plasma spraying), hardness of counter body (alumina, silicon nitride, and bearing steel) and normal load (10, 15, and 20 N) at dry sliding test on wear mass loss and coefficient of friction (responses). They proved that the hardness of counter body exerted the most significant influence on responses, followed by the type of deposition technology while normal load exerted the least influence. The smallest load (10 N), with the smallest hardness of the counter body (bearing steel) gave the lowest mass loss for the HVOF procedure.
Full factorial design was performed by Gil and Staia [79] while investigating corrosion resistance and porosity of HVOF sprayed Ni-based self-fluxing alloy coatings, varying spraying distance, powder feed rate, and fuel/oxygen ratio at three levels, while repeating all combinations of factor levels two times. They showed that all three factors exerted a significant influence on the result-corrosion current density and corrosion potential in 3.5% NaCl solution, but that only the spraying distance was of influence on porosity (for smaller distances, the porosity was even lower due to higher speed and higher temperature of melted particles). Optimal parameters were also defined: powder feed rate 60 g/min, spraying distance 380 mm, and fuel/oxygen ratio from 1.1 to 1.2.
Taguchi experimental design was applied for investigating HVOF and HVOLF (L-liquid) NiCrBSi alloy coatings deposited on the steel substrate [80]. Four factors were varied in slurry erosion examination: speed, concentration of slurry, impact angle, and size of abrasive particles. It was proven that the last-mentioned factor exerted no significant impact on wear rate. Table 3 summarizes the previously reviewed studies of Ni-based self-fluxing alloys, in which some optimization method was applied. PTAW-plasma-transferred arc welding; LC-laser cladding; FS-flame spraying; PS-plasma spraying; HVOF-high-velocity oxy fuel spraying; HVOLF-high-velocity oxy liquid fuel spraying.
Discussion
After reviewing the papers, it was noticed that, in addition to the statistical part, the microstructure of coatings was examined in most of the investigations. However, more attention should have been paid to the influence of deposition technology (or fusing) on the structure of the substrate. Evidence of a change in structure of coating, coating/substrate interface, and substrate is given due to high fusing temperature, depending on the type of substrate (and related thermal properties), coating thickness, but also on dimensions of the sample, which is proven in one of the previous works [7] of the authors of this paper. Figure 4 presents the flame spraying and simultaneous fusing, from which it is evident that there is a significant influence of heat on the substrate.
Discussion
After reviewing the papers, it was noticed that, in addition to the statistical part, the microstructure of coatings was examined in most of the investigations. However, more attention should have been paid to the influence of deposition technology (or fusing) on the structure of the substrate. Evidence of a change in structure of coating, coating/substrate interface, and substrate is given due to high fusing temperature, depending on the type of substrate (and related thermal properties), coating thickness, but also on dimensions of the sample, which is proven in one of the previous works [7] of the authors of this paper. Figure 4 presents the flame spraying and simultaneous fusing, from which it is evident that there is a significant influence of heat on the substrate.
It has already been stated that the fusing process is frequently performed after the flame and plasma spraying, at temperatures up to 1100 °C. However, some authors, especially for flame spraying, do not specify the method of fusing or do not mention it at all. In addition,, sometimes neither the type of material nor the dimensions of the substrate are stated. The authors very often do not even state the number of samples. Due to the nature of the thermal spraying technology (high cooling rates), the structure of the coatings is often unpredictable and it is good to have the replication principle fulfilled. In addition to this principle, the investigators should adhere to the other two principles of experimental design, which are randomization and blocking [17,18]. Therefore, the authors of this paper would like to suggest to the investigators to use more samples for the same combination of the levels of input variables. Likewise, randomization should be considered, as well as blocking, if necessary.
It was concluded that all authors have used engineering or process knowledge to analyze and interpret the results of the statistical analysis of experimental data, which is the single and best way to use DOE in any research. However, it has to be specified which design or designs of experiment were applied and why those types were chosen. For instance, given that RSM is a methodology which consists of several steps ( Figure 5) and is intended for optimization, researchers should distinguish between the terms RSM and response surface. The latter means the statistical model (i.e., the dependence of response on input variables) represented graphically by the surface. It has already been stated that the fusing process is frequently performed after the flame and plasma spraying, at temperatures up to 1100 • C. However, some authors, especially for flame spraying, do not specify the method of fusing or do not mention it at all. In addition" sometimes neither the type of material nor the dimensions of the substrate are stated.
The authors very often do not even state the number of samples. Due to the nature of the thermal spraying technology (high cooling rates), the structure of the coatings is often unpredictable and it is good to have the replication principle fulfilled. In addition to this principle, the investigators should adhere to the other two principles of experimental design, which are randomization and blocking [17,18]. Therefore, the authors of this paper would like to suggest to the investigators to use more samples for the same combination of the levels of input variables. Likewise, randomization should be considered, as well as blocking, if necessary.
It was concluded that all authors have used engineering or process knowledge to analyze and interpret the results of the statistical analysis of experimental data, which is the single and best way to use DOE in any research. However, it has to be specified which design or designs of experiment were applied and why those types were chosen. For instance, given that RSM is a methodology which consists of several steps ( Figure 5) and is intended for optimization, researchers should distinguish between the terms RSM and response surface. The latter means the statistical model (i.e., the dependence of response on input variables) represented graphically by the surface.
When it is about regression models, the authors of the present investigation would like to comment on this topic. By statistical processing of experimental data, using some statistical software, it is not enough just to derive a regression model. It needs to be checked, especially if used as input to single-and multi-objective optimization. In addition to analysis of variance (ANOVA) of the model, the diagnostics should include the following:
•
Normal probability plots of the residuals (or internally studentized residuals) to check for normal distribution ( Figure 6).
Plots of internally studentized residuals versus predicted values to check for constant error. •
Detailed discussion of externally studentized residuals to look for outliers, i.e., influential values (analysis of measure difference in fits and indicator of Cook's distance should be included).
•
Checking the coefficients of determination (high and equivalent values of R 2 adjusted and R 2 for prediction are desirable), as well as other ANOVA output (standard deviation, i.e., MSE-mean square error, predicted residual sum of squares, coefficient of variation, adequate precision, etc.).
The authors very often do not even state the number of samples. Due to the nature of the thermal spraying technology (high cooling rates), the structure of the coatings is often unpredictable and it is good to have the replication principle fulfilled. In addition to this principle, the investigators should adhere to the other two principles of experimental design, which are randomization and blocking [17,18]. Therefore, the authors of this paper would like to suggest to the investigators to use more samples for the same combination of the levels of input variables. Likewise, randomization should be considered, as well as blocking, if necessary.
It was concluded that all authors have used engineering or process knowledge to analyze and interpret the results of the statistical analysis of experimental data, which is the single and best way to use DOE in any research. However, it has to be specified which design or designs of experiment were applied and why those types were chosen. For instance, given that RSM is a methodology which consists of several steps ( Figure 5) and is intended for optimization, researchers should distinguish between the terms RSM and response surface. The latter means the statistical model (i.e., the dependence of response on input variables) represented graphically by the surface. When it is about regression models, the authors of the present investigation would like to comment on this topic. By statistical processing of experimental data, using some statistical software, it is not enough just to derive a regression model. It needs to be checked, especially if used as input to single-and multi-objective optimization. In addition to analysis of variance (ANOVA) of the model, the diagnostics should include the following: • Normal probability plots of the residuals (or internally studentized residuals) to check for normal distribution (Figure 6).
•
Plots of internally studentized residuals versus predicted values to check for constant error.
•
Detailed discussion of externally studentized residuals to look for outliers, i.e., influential values (analysis of measure difference in fits and indicator of Cook's distance should be included).
•
Checking the coefficients of determination (high and equivalent values of R 2 adjusted and R 2 for prediction are desirable), as well as other ANOVA output (standard deviation, i.e., MSE-mean square error, predicted residual sum of squares, coefficient of variation, adequate precision, etc.).
Moreover, significant interactions between factors of the experiment, as well as significance of factors, must be interpreted from the point of view of process knowledge. The possibility of model transformation must be considered, as well. Giving the experimental data in table form is much better than only graphical presentation of obtained results or not even presenting the data at all. The regression model diagnostics or important model statistics should exist, and the software used for statistical processing and analysis should be mentioned. Moreover, significant interactions between factors of the experiment, as well as significance of factors, must be interpreted from the point of view of process knowledge. The possibility of model transformation must be considered, as well.
Giving the experimental data in table form is much better than only graphical presentation of obtained results or not even presenting the data at all. The regression model diagnostics or important model statistics should exist, and the software used for statistical processing and analysis should be mentioned.
Conclusions
Due to the fact that Ni-based self-fluxing alloy coatings are often applied in protection from wear, a number of authors investigated wear. For the analysis of resistance to wear (sliding, abrasion, erosion, impact abrasion, erosion-corrosion), the response in most cases was the mass or volume loss or wear rate or coefficient of friction and the input variables represented a combination of the spraying parameters and the testing parameters (load, temperature, sliding distance or speed, size of abrasive testing particles, environment, samples' speed of revolution, impact angle, slurry concentration in investigation of slurry erosion, hardness, and number of revolutions of counter body) but also the properties of coatings (chemical composition of powders, presence of hard or solid lubricant particles, coating hardness, coatings with or without chrome added). Influence of the matrix hardness on wear was also studied.
In addition to wear, statistical approach was applied in the study of corrosion, adhesive strength, hardness, cracking resistance, surface roughness, and residual stresses. In treatments in which the effect on substrate is considerable (laser cladding, PTAW, and processes of remelting), geometric characteristics of the coating or heat-affected zone dependent on the treatment's parameters were a frequent response. Designed experiments or statistical methods are applied in the study of porosity and rolling contact fatigue, as well.
In addition to frequently varying the spraying parameters dependent on applied deposition technology, as well as powder feed rate and preheat temperature of the substrate, the spraying distance has also been often varied. Some authors defined optimal spraying distance and powder feed rate and it was proven that importance of these factors can depend also on type of response or testing and deposition technology used. For research connected with laser cladding or laser remelting, mostly, all authors varied laser power and scanning speed.
Based on the given review of the application of designed experiments and statistical methods in investigating thermally sprayed Ni-based self-fluxing alloy coatings, it can be concluded that a certain smaller number of authors combined designed experiments with artificial intelligence methods mainly aimed at obtaining optimal parameters of the applied deposition technology. It is obvious that there is an additional opportunity for research here. A paper was even found in which a combination of numerical modeling and experimental research was conducted, in which the results obtained by numerical modeling were compared with experimental values obtained by the application of designed experiment. Here, too, the possibility for further research exists. The designed experiments have also been applied with the aim to obtain optimal values of input variables for reaching minimal or maximal values of response but also to obtain regression models for demonstrating the dependence of response on input variables. In the research of some authors, designed experiments were not applied or specified, but the research results were demonstrated by statistical models, or some of the statistical methods were applied. | 15,218 | 2020-10-01T00:00:00.000 | [
"Materials Science",
"Engineering"
] |
Assessment Framework of Smart Shipyard Maturity Level via Data Envelopment Analysis
: The fourth industrial revolution (“Industry 4.0”) has caused an escalating need for smart technologies in manufacturing industries. Companies are examining various cutting-edge technologies to realize smart manufacturing and construct smart factories and are devoting efforts to improve their maturity level. However, productivity improvement is rarely achieved because of the large variety of new technologies and their wide range of applications; thus, elaborately setting improvement goals and plans are seldom accomplished. Fortunately, many researchers have presented guidelines for diagnosing the smartness maturity level and systematic directions to improve it, for the eventual improvement of productivity. However, most research has focused on mass production industries wherein the overall smartness maturity level is already high (e.g., high-level au-tomation). These studies thus have limited applicability to the shipbuilding industry, which is ba-sically a built-to-order industry. In this study, through a technical demand survey of the shipbuilding industry and an investigation of existing smart manufacturing and smart factories, the keywords of connectivity, automation, and intelligence were derived and based on these keywords, we developed a new diagnostic framework for smart shipyard maturity level assessment. The framework was applied to eight shipyards in South Korea to diagnose their smartness maturity level, and a data envelopment analysis (DEA) was performed to confirm the usefulness of the diagnosis results. By comparing the DEA models, the results with the smart level as an input represents the actual efficiency of shipyards better than the results of conventional models. H.C. Investigation, J.H.W.,
Introduction
Owing to changes in the global manufacturing environment, the shipbuilding industry currently faces challenges of survival and sustainability [1]. The global economic recession has induced shrinking international trade, while marine resource development projects are also decreasing or being canceled owing to declining oil prices, which are definitely unfavorable conditions for the shipbuilding and marine industry [2]. Analysts predict that it will be difficult to return to the early boom period of orders for ships and offshore plants, such as around 2010, which is referred to as the "super cycle" in the shipbuilding industry. Under these circumstances, shipyards worldwide are devoting various efforts to establish corporate strategies for the future. One such strategy is the active adoption of smart manufacturing technology, which numerous shipbuilding companies are considering [3,4].
Global manufacturing companies are expressing great interest in smart manufacturing and smart factories, with the goal of unmanned production through automation and connectivity among enterprise resources with the help of the IoT. The application of the smart factory concept is also accelerating with the adoption of Industry 4.0 in Germany. This concept is expanding with the combination of information and communications technology and automation solutions throughout the entire production process [5][6][7].
Although manufacturing companies seek to enhance the level of smart manufacturing through surveys on the latest technologies related to smart factories, adopting smart technologies often does not lead to practical effects (e.g., improved productivity). In fact, inelegantly applying smart technologies without sufficient accurate and practical analysis may adversely affect the existing production system. To solve this problem, a multidimensional and quantitative assessment of the smartness maturity level of manufacturing companies should be performed beforehand [8,9]. However, beyond simple products, it is necessary to consider both quantitative and qualitative criteria and interdependence in a complex manufacturing environment. Therefore, appropriate assessment of the manufacturing level of companies has long been an important issue. With rising interest in smart factories, researchers have proposed various definitions and criteria related to smart factories depending on the scale and type of the manufacturing company. Nevertheless, research on the assessment of smart factories is limited.
This growing emphasis on the importance of smart factories has led scholars to reexamine prior literature on the development of production systems and continuously publish new studies on the definition of a smart factory and the assessment of manufacturing systems. However, since most research has been focused on the general manufacturing industry, it has limited direct applicability to shipyards, which possess different industry characteristics. Basically, shipbuilding is a built-to-order industry. The products of the shipbuilding industry, i.e., ships-consists of several millions of similar but mostly unique intermediate products. At the early stages of production, the similarity among intermediate products is high enough to automate the processes, but the similarity plummets as the production progresses. Therefore, mostly, the automation level of shipyards are relatively low and disproportionally biased to earlier stages [10].
To evaluate smart shipyard and maturity level (SSML) assessment methods, the literature has been extensively surveyed. First, several studies examined smart factories in the manufacturing industry [6,7,11] Furthermore, studies on maturity level assessment of the manufacturing system can be largely divided into studies on evaluation of the manufacturing system in its current state and those on the evaluation of future manufacturing systems aimed at smart factories. In this regard, there have been several case studies on the evaluation of the manufacturing system in its current state [12][13][14][15][16]. In addition, since the advent of smart manufacturing owing to the fourth industrial revolution (4IR), maturity level assessment models for evaluating future manufacturing systems have been studied since 2013 [17][18][19][20]. Detailed descriptions of the literature on the definition of smart factories and on manufacturing system assessment are shown in Appendix A.
In South Korea, the Korea Production System (KPS) was developed, which is suitable for the South Korean manufacturing environment across representative manufacturing sectors such as automotive and consumer goods. [21]. Furthermore, following the development of KPS, to successfully promote the spread of smart factories, the Smart Manufacturing Innovation Planning Division of the Bureau of small and medium-sized enterprises (SMEs) developed a diagnostic tool that can represent plans for smart factory construction, with the goal of objectively diagnosing and assessing the smart maturity level of the manufacturing industry [22]. In this study, they derived the criteria and modules for smart factory assessment from the framework for the smart factory operation system (vision, goal, enterprise, factory, machine, and control). In addition, a questionnaire for the assessment items comprising each module was defined according to the maturity level definitions. Detailed configurations of KPS and smart factory assessment modules are shown in Appendix B.
The smart factory assessment module is a diagnostic tool suitable for the general machinery sector and is actively used by South Korean manufacturing companies [22]. Accordingly, this study also attempted to assess South Korea's large and medium-sized shipbuilding companies using this smart factory assessment tool in the initial stage. However, the composition and contents of the assessment items defined in the existing model did not reflect the characteristics of the shipbuilding industry and differed from the shipbuilding production system, thus, degrading the reliability of the assessment results.
Concerning multivariate analysis, mathematical analysis techniques such as analytic hierarchy process (AHP), data envelopment analysis (DEA), and multivariate regression analysis are widely used for the assessment of multiple targets derived from various input variables. A meaningful productivity predictor was proposed based on a regression analysis of various indicators related to shipyard productivity [23]. The competitiveness of shipyards was analyzed through DEA, using capacity, technology level, and the industrial environment of various shipyards as input and productivity and building times as output [24]. Their studies extended beyond a qualitative analysis; through DEA, they expressed the relative levels of the most advanced shipyards and those lagging behind them, thereby presenting directions for improvement and quantitative improvement levels. A shipyard efficiency analysis using DEA was also introduced [25]. Turnover/cost values were derived by combining various shipyard production indicators to compare the actual competitiveness of shipyards [26]. Moreover, an assessment methodology for the shipyard block assembly process was developed using process mining and DEA, and they described a practical case applying the methodology to shipyards [27]. In addition, shipyard production plans were assessed using the AHP method considering sales, dock turnover, and quay load [28]. Most recently, a comparative analysis was performed for assessing the productivity of 21 shipyards in South Korea, China, and Japan using DEA and the metafrontier framework [29]. As such, researchers have performed a variety of studies on shipyard competitiveness and productivity, with DEA being applied the most.
In this study, we propose a SSML assessment framework that considers ship production characteristics and diverse production environments. To this end, we devised a modified assessment framework appropriate for the SSML assessment based on the categories of the smart factory assessment framework proposed by Lee et al. [22]. For this purpose, the SSML assessment framework was defined through expert surveys on South Korean shipbuilders of various sizes and technology levels and reflected in the composition of the detailed assessment items. Furthermore, we applied the developed SSML assessment framework to real shipyards and their subcontractors. To apply the proposed framework, this study adopted DEA techniques [24,30] to analyze productivity in the shipyard industry. Using DEA, we built a model that set the maturity level of the shipyard and its subcontractors as input, and sales and order quantity as output. The model confirmed that the smartness maturity level can serve as useful information for assessing the capabilities of companies and deriving quantitative improvement indicators.
The remainder of this study is structured as follows. In Section 2, we present how to develop the SSML and, in Section 3, we provide the results and analysis of the assessment from the interview. In Section 4, we introduce DEA and describe the method of quantitative analysis using DEA. In Section 5, we present the results and discuss the practical contributions of this study. In Section 6, we draw conclusions and suggest future research.
Development of Smart Shipyard Maturity Levels
While these previous maturity models present meaningful diagnostic criteria, they either lack specificity or are difficult to apply to diagnosing the SSML. As incorrect assessment criteria lead to incorrect assessment results, smartness maturity levels that reflect the characteristics of shipbuilding production systems must be defined.
With few studies on the assessment of shipyards, this study is the first step for developing a smart maturity level assessment system for shipyards. Technical demand surveys were conducted with large and medium-sized shipyards in South Korea, module manufacturing subcontractors, and related research institutes. The surveys were conducted by e-mail on May 2018, with the shipbuilding industry partners in South Korea including large shipyards, subcontractors, and research institutes. Through these online and offline demand surveys, we identified approximately 450 technology demand surveys from smart shipyards, which were categorized into five areas, as shown in Figure 1. Next, an industry-academia-research expert group was formed to devise a technical roadmap for realizing smart shipyards, see Figure 2, based on the technical demand survey results. This roadmap comprised a bottom-up process for selecting four major fields from the technology demands of the 450 candidates, and a top-down process for classifying the detailed tasks through the expert group. The technical roadmap was largely classified into infrastructure technology, including IoT, big data, process automation for unmanned production and logistics activities, intelligent technology for an array of production management tasks, and production design automation technology. Next, to prepare the standard for the SSML from the technical roadmap, we used the concept revealed from the existing research focusing on the automation concept It is difficult to define the levels for complex production systems such as shipyards because existing level definitions for smart factories are defined on a single-level scale. The concept of automation was examined from a human-centered perspective rather than a physical device [31]. Accordingly, we defined the types of automation as follows: 1. Control automation assists humans in the guiding process of executing the task and the machine movement through dangerous tasks. Control automation plays the role of an observer of the whole subsystem. 2. Management automation allows humans to exercise demands oriented to technological actions and activities, and also a strategic point of the automation process. 3. Information automation, a type of system that is changing very rapidly, provides the system with information about the progress and the execution of certain tasks.
We used these three types of automation, i.e., physical, intellectual, and information transaction, as keywords for smart shipyards (see Table 1). First, physical automation corresponds to control automation and automation of human physical labor. Next, intellectual automation refers to automation of human knowledge labor and corresponds to the management automation. Finally, connectivity refers to automation of information transfer, and corresponds to information automation [31]. Additionally, the assessment criteria for each maturity factor and level, together with the specific integrated form of smart shipyard concerning each level, are shown in Appendix C. On the basis of this definition of SSML and on the smart factory assessment framework [22], an assessment framework for SSML was developed. In this process, the number of inquiry items in the existing model increased from 46 (from [22]) to 61 (assessment framework for SSML) to reflect the characteristics of the shipbuilding industry. In particular, assessment items for the logistics operation and information system modules were added and modified (Table 2). Furthermore, Figure 3 shows the structure of the criteria, modules, and assessment items of the proposed assessment framework. This diagnosis framework was divided into the following four criteria groups: leadership and strategy, process, system and automation, and performance. In addition, each criteria group consisted of submodules; in particular, the process criteria group included product development, production planning, process management, quality management, facility management, and logistics operation modules in consideration of the shipbuilding process. Each module included assessment items, and based on those, the smart shipyard maturity level was diagnosed. The composition of a comprehensive system for production planning and service levels concerning information systems
Methodology for Assessment of Smart Shipyard and Maturity Level (SSML)
We conducted an assessment of shipyards to verify the developed assessment framework. Regarding the assessment procedure, first, the assessment material was delivered to 15 companies, including large (L1~L5), mid-sized shipyards (M1~M3), as well as subcontractors such as hull block assembly (B1 and B2), outfitting material manufacturers (O1 and O2), and steel fabrication companies (P1 and P2), and the person-in-charge of the respective module at each company completed the assessment material through a self-assessment. As shown in Figure 3, since there were 10 evaluation modules, 10 employees of each company participated in the assessment. In order to eliminate personal bias, two or three consulting experts who had over 10 years of experience in shipbuilding production area visited each company afterwards and corrected those self-assessment scores through in-depth interviews with the person-in-charge of the respective module and site inspections. Then, reports were written on the interview, site inspection results, and corrected diagnoses and delivered to each company, after which feedback was received. The assessment results were converted into a score of 5 points and Table 3 shows the results of SSML of each assessment module.
SSML Assessment Results and Discussion
In contrast to general business consulting, the assessment results were expressed as scores from a sophisticated assessment that reflects the widely known scale and qualitative level of companies. Furthermore, they are meaningful as much as this is the first assessment of detailed production factors in the field of shipbuilding production. Figure 4 shows the average score for each module of the companies investigated. Production automation, the most noteworthy module, showed the lowest assessment at 1.9 (approximately 40 points based on 100 points) owing to the high dependence on workforce in the shipbuilding industry. Most shipyard and subcontractor processes are manual processes. All construction can be performed manually, excluding processing and some assembly processes in large and mid-sized shipyards. Considering the weight of the number of hours by process, large shipyards also have very low automation rates, at approximately 30% for processing and 10% for assembly. Next, the levels of quality management, facility management, and logistics operations were diagnosed as low. Concerning quality management, although domestic shipyards are somewhat competitive internationally [32], this module was assessed to be low because the assessment focused on the computerization aspect of data collection, processing, and sharing, rather than the level of quality itself. However, the numerous complaints in the on-site interviews from quality management employees about the inadequate data management systems suggests that, in addition to the quality management capabilities of domestic shipyards, the informatization systems that support quality management must be improved. The facility management assessment signifies inadequate management concerning smart maturity of large transport equipment (gantry cranes, transporters, etc.) and production equipment (cutting, welding, painting equipment, etc.). Advanced technologies such as connectivity and predictive maintenance, which are pursued in smart production, were not adopted, and most maintenance procedures consisted of responding to problems after they occurred. Hence, there is a need for advanced facility management by applying predictive maintenance and IoT connectivity technologies, which automatically collect equipment information in real time. As one of the main targets of shipyard management, logistics was recognized as a component of production that enables the smooth flow of production activities by connecting processes, rather than as a simple transport activity. Nevertheless, shipyards remain in the process of adopting logistics technology for tracking work in process (steel plate, hull block, outfitting module, etc.) items; therefore, from the perspective of smart production that pursues connectivity and automation in transport, the low score for this module is reasonable. Aside from these modules, the scores for product planning, production planning, process management, and the information system (approximately 2.7-2.8) were higher than the overall average smart maturity level.
Management's proactiveness concerning smart technology, in terms of leadership and strategy and performance measurement, was scored relatively high at 2.9 and 3.0, respectively. However, as these two modules are considerably closer to qualitative assessment than other modules that can be quantitatively diagnosed (systems, facilities, informatization, etc.), they can be regarded as assessment items not supported by concrete evidence. Therefore, since the weights among the modules must be considered, future studies must apply the AHP technique to derive more reliable assessment results.
Next, we conducted a clustering analysis of each company based on the maturity level scores. Excluding leadership and strategy and performance measurement, which directly impact the production level, the average scores of the "process" and "system and automation" groups based on the criteria were set as the x-and y-axes, respectively, and Figure 5. In Figure 5, the size of each circle indicates a company's relative sales in 2019. As the low tide of the international shipbuilding market continued in 2019, the relative sales of each company indicated in the graph are not proportional to their production capacity. Therefore, although M2/M3 and M1/M4 are classified as mid-sized shipyards, the difference in actual production capacity is very large. Nevertheless, the scale of sales shown in Figure 5 are similar owing to the influence of market conditions; therefore, they should be considered only for the purpose of classifying company type. This is shown to classify large and mid-sized enterprises/SMEs. The large L1-L5 enterprises formed a relatively higher cluster, while the level of M2 and M3 companies among the mid-sized enterprises was similar to that of large shipyards. Although M1 and M4 are also classified as mid-sized shipyards along with M2 and M3 because the sizes of M2 and M3 are close to large, the difference in size is also reflected in the production level. The outfitting manufacturers (O1, O2) and block manufacturers (B1, B2), as well as mid-sized shipyards M1 and M4, formed a cluster separate from the large enterprise cluster. P1 and P2 located between these two clusters reflect the characteristics of companies that form curved hull plates, which showed rather exceptional results. These results suggest that, as basic research on the automation of curved hull plate forming has recently matured to some degree, curved hull plate forming companies have also attempted to replace operator-dependent tasks with automated machines, thus, resulting in a movement toward automation reflected in the smart maturity level.
By analyzing the company type and SSML score as above, we could quantitatively analyze the smart maturity level of large-and mid-sized companies in the South Korean shipbuilding industry. We confirmed that the developed assessment framework for the SSML can reasonably quantify the types and levels of shipbuilding companies.
Data Envelopment Analysis (DEA) Method
The proposed DEA method measures efficiency based on linear programming in the decision-making process [33]. This method is mainly applied when it is difficult to identify and compare direct relationships among multiple inputs and outputs and can be used to develop a more suitable decision-making model than AHP in environments with insufficient information. AHP and DEA methods have been applied to manufacturing supply [34]. According to the findings, the AHP technique can provide detailed and gradual decision-making results through pairwise comparison in cases with large amounts of information, whereas the DEA method can provide effective decision-making strategies for situations with insufficient decision-making factors in a new environment. Accordingly, this study focused on a situation with a lack of specific grounds and information for smart shipyard technology and directions and noted very fast and uncertain technological change in the shipyard industry, unlike existing manufacturing technologies. Therefore, rather than AHP, which is time-consuming in the development process, DEA was applied using the assessment results of the SSML as input, with the goal of verifying whether it could be useful for calculating the efficiency of shipyards and deriving quantitative improvement indicators.
This method calculates the relative efficiency of assessment targets, where multiple inputs and outputs are considered. It derives the most efficient DMUs from all DMUs to be assessed; consequently, the relative efficiency of each DMU is calculated using linear programming. DEA models are largely categorized into models that assume constant returns to scale (CRS) and those that assume variable returns to scale (VRS). The CRS model assumes that the relationship between input and output is the same at a constant rate, regardless of scale, and was used when the DEA methodology was first proposed. The CRS model is also referred to as the Charnes-Cooper-Rhodes (CCR) model, after the first proposed model. The VRS model [30] relaxes the assumption of CRS in the CCR model and is also referred to as the Banker-Charnes-Cooper (BCC) model.
DEA models can also be categorized into input-oriented and output-oriented models according to their orientation to input or output. The input-oriented model seeks to minimize input with a fixed output, whereas the output-oriented model seeks to maximize output with a fixed input.
DEA models are based on linear programming and can be explained by the following equations: First, assume that n DMUs exist. If (k = 1, …, n) means that m inputs (i = 1, …, m; j = 1, …, n) are input to output s outputs (r = 1, …, s; j = 1, …, n), then, the efficiency of the th observation is obtained through the linear programming solution of Equations (1) and (2) assuming an output-oriented CCR model. In the equation, is the efficiency value, and and are slack variables for the input and output, respectively. If the value of * obtained as the solution to this equation is 1 and both slack variables are 0, then the DMU is an efficient (efficiency 100%) DMU. Equations (1) and (2) are as follows: . .
As the BCC model assumes VRS, a constraint is added such that the sum of is 1. Accordingly, the output-oriented BCC model can be expressed as in Equations (3) and (4) as follows: . .
If the efficiency value of the CCR model is * and that of the BCC model is * , then, the constant relationship * ≤ * is established, and the difference in efficiency between the two models originates from whether the scale is optimal. This difference is the SE, and the following relationship is established: If SE is 1, then, it is in a CRS state and there is no inefficiency owing to scale; if SE is less than 1, then, it is in an increasing or decreasing returns to scale state, signifying that there is inefficiency owing to scale. The efficiency of the CCR model ( * ) is referred to as the technical efficiency, while the efficiency of the BCC model ( * ) is referred to as pure technical efficiency to emphasize that inefficiency owing to scale is excluded.
Quantitative Analysis Using DEA
Next, using the DEA method, we conducted a quantitative analysis, compared the companies, and presented target levels of quantitative improvement for relatively lowlevel companies. DEA is an analysis method that measures the relative efficiency of companies with multiple input and output factors. It measures the relative efficiency between the same groups and provides information on benchmarking targets to improve efficiency for those that appear inefficient. The DEA technique is useful for the following problems.
1. It is useful when there are many inputs and outputs, but it is difficult to integrate them into the single index in an appropriate way. 2. DEA provides a basis for benchmarking target that should be investigated to improve efficiency. 3. DEA can simultaneously consider various input and output factors with different units of measurement.
To measure efficiency using DEA, decision-making units (DMUs) must be set. DEA presupposes the homogeneity of the analysis target, requiring individual DMUs to perform similar activities to produce products that can be compared [35]. Furthermore, similar resources or capital must be input in all DMUs, and performance must not be influenced by external factors. Accordingly, the DMUs in this study, consisted of shipyards focusing on new shipbuilding. However, shipyards for which output variable data (new ship construction) could not be secured were excluded from the analysis.
The selection of the input and output variables is important for ensuring the reliability of the DEA results. That is, it is necessary to select input and output variables that can accurately reflect the objectives, targets, and production environment of the shipyard. Our objective was to verify whether the SSML assessment result is an indicator of shipyard efficiency. As such, a DEA model was built using the number of employees and docks as inputs and the new ship construction as the output.
Furthermore, as described above, the DEA model was categorized into a CCR model (for Charnes, Cooper, and Rhodes [33]) and BCC model (for Banker, Charnes, and Cooper [30]) based on the assumption of the effect of scale, and also categorized into an inputoriented (minimum input) model and output-oriented (maximum output) model, depending on the purpose of measuring efficiency. Regarding the selection of the model, since new ship construction is greatly affected by external factors, the input-oriented (minimum input) model was selected, which minimizes the input for a fixed output. The commercial software Frontier Analyst was used for the DEA.
Next, two models were defined as shown in Table 4, to perform the DEA. However, as the input and output of the DEA model are proportionally correlated, for the SSML score (the input variable), the proposed model uses the value of the perfect score (5.0) minus the score of each factor. This was a given correlation as an input factor of the DEA model, in which the lower the maturity level score, the closer it is to the advanced level. That is, a high maturity level indicates a relatively small input factor value, meaning that the output can be achieved with less effort. Thus, the maturity level was analyzed by subtracting the projection value derived via DEA from the perfect score (5.0). Owing to insufficient data, DMUs could not be performed for all companies in Table 3. However, a DEA was conducted for shipyards with available data on output Y1 (L1-L5, M2, M3, and M4) common to the two models and X1 and X2 of the traditional model in Table 4.
There are considerations concerning the timing of the input and output data. As shipyard construction volume greatly varies with the time period, large gaps in the potential production volume of a shipyard may arise when using the proposed model's input data X1-X2 and Y1 of when the maturity level assessment was performed in 2009. Moreover, data on the shipyard's production capacity are not explicitly disclosed or available. Accordingly, for the data of X1-X2 and Y1, we considered the largest volume of new ship construction during the last 10 years and the number of employees and docks/berths in that year. Consequently, the 2019 data were selected. Although applying these conditions will lead to mismatched timing in the output and input data, we ignored it because our objective was to compare the results of DEA, which uses physical conditions (workforce, resources, etc.) as input, with the results of the proposed method, which uses the maturity level as input, rather than present quantitative information on each shipyard through precise assessment results. Furthermore, as the speed of change in shipyards is slow, it was judged that the level assessed in 2019 had not dramatically changed over the last 10 years.
Before describing the analysis, the following assumptions were made: If the objective reliability of the relationship between the input and outputs considered in Table 4 and the DEA method is secured, then the level (or efficiency) of each shipyard will be determined by the DEA results. However, we aimed to identify which input/output and DEA model best expressed the phenomenon (shipyard level); the judgment criteria for the suitability of the analysis results were defined as the shipyard level, which were generally known as follows. For the assumptions of the relative levels of L1-L5 and M2-M4, they are classified based on the global ranking of each shipyard. First, L1-L3 are within the top five shipyards worldwide, with little change in ranking over the past 20 years. Thus, their scale, as indicated by sales and orders, and also their overall production level is among the best worldwide. Next, L4 and L5 are within the top 10 shipyards worldwide; although, the scale is somewhat smaller than L1-L3, this does not necessarily indicate a larger difference in production level. Next, M2 and M3 are shipyards in the top 10-50 worldwide, with relatively high-ranking volatility. Finally, M4 corresponds to shipyards ranked 50-100. However, as the difference in order and construction volume decreases at lower ranks, the difference in SSML does not increase as much as the difference in rank. Accordingly, L1-L3 is defined as the large group, L4 and L5 as the mid-sized-large group, M2 and M3 as the mid-sized group, and M4 as the small group, which are assumed to be the guidelines for analyzing the DEA results. According to the previous study, if the skill level of a large shipyard is set to 100, the level of a medium-sized shipyard is 75-85, and that of a small shipyard is approximately 50 [36]. Therefore, this quantitative guideline is used as comparative data to examine the efficiency feasibility of each shipyard through DEA analysis.
Results of the Data Envelopment Analysis
DEA was performed for Case 1 of a single model consisting of L1-L5 and M2-M4 as DMUs and Case 2 of split models with L1-L5 and M2-M4 separated.
Case 1
In Case 1, L1-L5 and M2-M4 were configured as one model. Figure 6 shows the efficiency and the scale efficiency (SE) results derived by applying the CCR and BCC analysis methods concerning the traditional model and proposed model. In Figure 6 (raw data is Table A8), SE values below 1 were derived. If the SE is less than 1, the DMUs are in a state of increasing or decreasing returns to scale states, signifying that there is inefficiency owing to scale. Next, the results were derived from the CCR and BCC methods using the traditional and proposed models. Figure 6 shows the analysis results of the BCC and CCR methods, under traditional analysis conditions using the number of employees and docks as input variables. First, concerning the BCC method, the efficiencies of all shipyards, excluding M2 in BCC and M2 and M4 in CCR, were 100% or close to 100%, confirming that there was no discriminatory power between the shipyards. Furthermore, concerning both BCC and CCR, L1 and M2 was lower than M3 and M4, although it ranked higher than both in reality; hence, they did not meet the guidelines assumed in Section 4.
Next, Figure 6b shows the analysis results from BCC and CCR, under the analysis condition using two previously analyzed SSML categories as the input variables; specifically, the average of the perfect score (5.0) of each module minus the score of each factor.
In Figure 6b, the CCR analysis results under the proposed model show efficiencies of 21.56%, 7.95%, and 4.57% for M2-M4, respectively, thus exhibiting differentiation with L1-L5. However, the difference became excessively large, resulting in practically impossible values for the projections of process and system and automation, which are inputs. Therefore, the analysis results are unreasonable. On the contrary, the BCC results, in Figure 6b, show the analysis results from the BCC method under the proposed model. All shipyard efficiency results using the BCC method under the proposed model satisfied the assumed guidelines in Section 4. L1-L3 was at 100% and was located in the best practice line, followed by L4 and L5 at approximately 80%, showing a difference of approximately 20% with the large group. Next, M2 and M3 showed values in the upper 70%, slightly less than the mid-sized and large groups; and M4, belonging to the small group, was approximately 60%. Thus, in the proposed model, the BCC analysis presents appropriate effi-ciency and projection values considering the qualitative level of each shipyard. Concerning the construction volume projections (as marked in red in Table A11), however, the values of M3 and M4 were unrealistically overestimated. This problem likely occurred because the difference in construction values from other shipyards was too large, despite being an input-oriented model. Accordingly, in the DEA analysis of Case 2, L1-L5 (large group) and M2-M4 (small and mid-sized group) were divided into separate models.
Case 2
Owing to the large difference in scale between the large-, small-, and mid-sized groups, the construction volume projections calculated for the small-and mid-sized shipyards were overestimated. Hence, in Case 2, L1-L5 (large group) and M2-M4 (small and mid-sized groups) were divided into separate models for the analysis. Since BCC showed more reasonable results than CCR in the analysis of Case 1, only the results under the proposed model were addressed in Case 2. For reference, Table A14 and Table A15 in Appendix D show the detailed results under the traditional model for Case 2. Figure 7 summarizes the efficiency and SE results. According to the results under the proposed model of Figure 7b,d, there was no change in L1-L5, which belongs to the large group, with only a change in M2-M4. That is, since the model for the small group was separated considering the difference in construction size, the large group showed no change from the existing results; and M2-M4, which were separated from the construction of large shipyards into a group of small shipyards, were calculated independently, thus adjusting the efficiency. Comparing the efficiency of CCR and BCC using the proposed model in the M2-M4 analysis, BCC (100%, 100%, and 80.22%) better reflects the shipyard guidelines than CCR (100%, 40.55%, and 21.20%). In other words, since M2 and M3 correspond to mid-sized shipyards with similar scale and technology level and M4 corresponds to small shipyards, BCC exhibited more reasonable results than CCR.
Additionally, Table A16 in Appendix D shows the DEA results of proposed approach and explains the detailed analysis results from the BCC methods. As shown in Table A16, the process and system and automation levels of L4 and L5 (large shipyards) must be improved by approximately 11-16% to reach the levels of L1-L3. As for M2-M4, whereas M2 and M3 are at 100% efficiency within that group, M4 shipyards must improve their process level by 20% and system and automation by 27%.
As a result of DEA analysis with respect to various factors and cases as compared with the existing DEA technique which uses physical elements such as the number of employees and docks as input, the proposed method, which uses the SSML as input, derived more reasonable results for the efficiency and projection of actual shipyards.
Conclusions
To develop a framework for diagnosing the smart maturity level of the shipbuilding industry, this study analyzed existing research on smart manufacturing, smart factories, and maturity models. Technology demand surveys were also performed to reflect the characteristics of the shipbuilding industry, and a technical roadmap for realizing smart shipyards that reflected the opinions of an expert group was proposed. The SSML assessment framework developed through this process defined five maturity levels for each of the following keywords: connectivity, automation, and intelligence. Furthermore, based on the defined levels, a diagnostic tool comprising 61 items for four criteria and 10 modules was developed. While the structure presented in prior research was used for the criteria and modules of the diagnostic tool, the 61 detailed inquiry items were reconfigured reflecting the characteristics of shipyards. This developed framework was used to diagnose large-and medium-sized shipyards and subcontractors in South Korea, after which the results were analyzed. Automation was assessed the lowest in the maturity level of shipbuilding-related companies, and it was confirmed that those companies could be divided into groups through a bubble chart analysis using process and system/automation criteria as the two axes.
Next, the assessment of SSML was used to conduct a DEA, which was capable of quantitative analysis. Using the SSML as the input variable, DEA can derive the efficiency levels of the subject companies and confirm the level of improvement required to reach 100% efficiency for each module through quantitative indicators. In addition, for the same companies, we performed a comparative analysis between the traditional DEA model, which used the number of employees, scale of facilities, etc. as input variables; and the other DEA model, which used the SSML as the input variable. The results demonstrated that the model using the maturity level as the input variable derived more reliable results that were well matched with a previous survey [36].
However, for DEA to be more pertinent, the number of DMUs must be at least two to three times greater than the sum of the number of input and output variables [37]. As the assessment was based on a rather insufficient number of DMUs (companies to be diagnosed), further investigations with more DMUs are strongly recommended, including not only South Korean shipyards but also shipyards of similar scales in China, Japan, and Europe.
Data Availability Statement:
The data presented in this study is in Table 3 and Appendices A-D.
Conflicts of Interest:
The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.
Appendix A. Smart Factory Definition and Research on Manufacturing System Assessment
To evaluate SSML assessment methods for the smart level definition, the literature on the smart factory definition required was extensively surveyed. First, for a smart factory, one of the biggest topics in the manufacturing industry, similar concepts were defined in the following research cases: A smart factory was defined as a context-aware manufacturing environment that can respond to disruptions in real-time production using distributed information and communication structures to optimally manage production processes, and as the model of next-generation factories in an era of ubiquitous computing technology [6]. It was defined as a factory of the future and factory-of-things "composed of smart objects which interact based on semantic services," and emphasized that, rather than a hierarchy in the traditional sense, the objects will self-organize to fulfill certain tasks [7]. SmartFactoryKL which is a demonstration and research test bed for smart factories was introduced in that study conducted by. The German Research Center for Artificial Intelligence (Deutsches Forschungszentrum für Künstliche Intelligenz, DFKI). The United States National Institute of Standards and Technology (NIST) defines smart manufacturing as a fully integrated cooperative manufacturing system that responds in real time to the changing demands and conditions of factories, supply chains, and customer needs, and emphasizes integration not only within the factory but also with supply chains and customers [11].
Studies on maturity level assessment of the manufacturing system can be largely divided into studies on evaluation of the manufacturing system in the current state and those on the evaluation of future manufacturing systems aimed at smart factories. First, several case studies evaluated the manufacturing system in its current state. Existing manufacturing system maturity level methodologies related to manufacturing SMEs was examined and an improved assessment tool was proposed [12]. In a project with the Mechanical Engineering Industry Association in Germany (Verband Deutscher Maschinenund Anlagenbau), the "Industry 4.0 Readiness" model was proposed [13]. This assessment model comprised six dimensions and 18 fields; the six dimensions consist of four dimensions (smart factory, smart product, smart operation, and data-driven services) in addition to "strategy and organization" and "employees." Maturity level assessment models provided by accredited institutions include the following: The Capability Maturity Model Integration (CMMI) model developed by the Software Engineering Institute (SEI) starts from a five level maturity model of software development and serves as the basis for various process maturity models, such as purchasing, products, and human resources [14]. Concerning manufacturing competitiveness, the Manufacturing Enterprise Systems (MES) Maturity Profile, developed by the MES Association, presents the maturity levels for MES, a system for optimizing production from product order to delivery [15]. The Business Process Maturity Model (BPMM) of the Ob-ject Management Group (OMG) presents an organization's processes as a five-level maturity model. The BPMM was developed as a diagnostic tool that substitutes critical success factors (information technology and systems, culture, responsibility, methodology, performance, etc.) and the perspectives of connection, design, execution, control, and improvement in CMMI [16]. Table A1 shows the maturity levels of SEI, MES Association, and OMG. Managed work unit mgmt.
Standardized process mgmt.
Since the advent of smart manufacturing owing to 4IR, maturity level assessment models related to 4IR have been studied since 2013, as shown in Table A2.
The Connected Enterprise Maturity Model developed by Rockwell Automation [17] presents enterprise maturity at five levels (assessment, secure and upgraded network and controls, defined and organized working data capital, analytics, and collaboration), but does not present detailed assessment items and a development process. The Reifegradmodell Industrie 4.0 (4IR), jointly developed by Mechatronics Cluster and Upper Austria University of Applied Sciences [18], presents three areas of data, intelligence, and digital transformation, 13 assessment items, and uses a 10-point-scale maturity level assessment method. However, the maturity model and detailed assessment content are inadequate. The Industry 4.0 Self Assessment model developed by pwc [19] derived six areas related to 4IR competencies (digital business models and customer access; digitization of product and service offerings; digitization and integration of vertical and horizontal value chains; data and analytics as core capability; agile IT architecture, compliance, security, legal, and tax; and organization, employees, and digital culture) and presented a four-level maturity model; however, only some evaluation items and no detailed development process were presented. The smart manufacturing system readiness level model under development by researchers at NIST diagnoses the maturity level for four areas, i.e., organizational maturity, IT maturity, performance management maturity, and information connectivity maturity [20]. However, based on an improvement activity model and factory design using IDEF0 (Integration Definition 0 [38]), it assesses the readiness level and focuses on improvements in the information system sector, making it somewhat inadequate for comprehensive production system diagnoses.
Appendix B. Korea Production System and Smart Factory Assessment Modules
KPS, a part of the Korean government's strategy for enhancing the productivity of manufacturing companies, is a tool that integrates various manufacturing innovation methodologies to establish a standard production system appropriate for the manufacturing capabilities and industrial culture of Korean manufacturing companies and to improve productivity.
Following the development of KPS, to successfully promote the spread of smart factories, the Smart Manufacturing Innovation Planning Division of the Bureau of SMEs developed a diagnostic tool that can present plans for smart factory construction, with the goal of objectively diagnosing and assessing the manufacturing industry's smart maturity level [25] (see Table A3). This diagnostic tool follows ISO 9001:2015 (management system), IEC 62264 (manufacturing operation system), ISO 22400, and SCOR (KPI); and was developed for certifying factory operation systems and designed to be linked with in-house enterprise certification systems, reflecting the culture and characteristics of the Korean manufacturing industry (Samsung Electronics, Hyundai Motor Company, POSCO, etc.). Moreover, as shown in Figure A1, a framework for smart factory operation was constructed, and smart factory assessment modules were derived. The criteria and modules for smart factory assessment (Table A4) from the framework for the smart factory operation system was derived, as shown in Figure A1 (vision, goal, enterprise, factory, and machine/control) [22]. In addition, a questionnaire for the assessment items comprising each module was defined according to the maturity level definitions in Table A5. Table A7. Description of each level of shipyard.
Level 5 shipyard
Artificial intelligence that transcends human thinking; that is, not only information generation, but also the design and execution of the information generating process itself is performed by a computer. Therefore, no human intervention is required for knowledge activities through information generation and information analysis. Data and information between human and objects are shared by IoT networks that are deployed across the enterprise.
In addition to the level of automation of Level 4, the production activity is also embedded in the production resource, and the occurrence of abnormal situations is autonomously managed by predictive preservation technology synchronized with production activities.
Level 4 shipyard
Information is generated by artificial intelligence algorithms at the level of human thinking, and data and generated information are automatically shared between computers and IoT devices according to a predefined work process without humans or devices.
Production is automatically performed by autonomous facilities.
In addition, product movement is also performed by automated transportation facilities. However, if a failure occurs in the facility or a work abnormality occurs, an abnormal signal is immediately transmitted to the work manager and the abnormal situation is managed.
Level 3 shipyard
When a person enters data in a computer equipped with an engineering algorithm, information is generated by the algorithm, and the generated information is shared between departments by a person using a wireless device, and production data are processed through a wireless device.
Unit production work is performed by being delivered to an independent automation facility on-site.
However, the scope of automation is limited to the unit process (cutting, grain processing, welding of sub-assembly, etc.), and if a failure in the automation facility or a work abnormality occurs, the abnormal situation is managed by the waiting work manager.
Level 2 shipyard
Data produced by human experience and ability is stored in a wired OA device, and the stored data are transferred in a manner of inquiry/copy/modification/distribution using wired OA devices and programs.
Data for production is delivered to the operator through the OA device, and the data received by the operator is loaded onto the machine or the machine is operated according to the received work instruction to perform production activities.
Level 1 shipyard
A shipyard where production data are created based on human experience and ability, and the generated data are delivered through direct human conversation instructions and paper documents without separate storage, and production activities are performed by direct human manual work. In Tables A11 and A12, the "data," "projection," and difference rows for "process" and "system and automation" indicate the input data, quantitative value to reach 100% efficiency, and ratio of the projection and data, respectively. Two values are shown for each data, projection, and difference by the DMU in the proposed model. The value on the left is the perfect SSML score of the perfect score, which is 5.0 minus assessment value and the projected value, and the value on the right is converted from the perfect score again. | 11,128.6 | 2021-02-11T00:00:00.000 | [
"Computer Science",
"Engineering",
"Environmental Science"
] |
Factor Xa inhibitor, edoxaban ameliorates renal injury after subtotal nephrectomy by reducing epithelial‐mesenchymal transition and inflammatory response
Abstract Chronic kidney disease (CKD) is an increasing and life‐threatening disease worldwide. Recent evidence indicates that blood coagulation factors promote renal dysfunction in CKD patients. Activated factor X (FXa) inhibitors are safe and first‐line drugs for the prevention of thrombosis in patients with atrial fibrillation. Here, we investigated the therapeutic effects of edoxaban on CKD using the mouse 5/6 nephrectomy model. Eight‐week‐old wild‐type mice were subjected to 5/6 nephrectomy surgery and randomly assigned to two groups, edoxaban or vehicle admixture diet. Edoxaban treatment led to reduction of urinary albumin excretion and plasma UN levels compared with vehicle group, which was accompanied with reduced glomerular cross‐sectional area and cell number. Edoxaban treatment also attenuated fibrinogen positive area in the remnant kidneys after subtotal nephrectomy. Moreover, edoxaban treatment resulted in attenuated tubulointerstitial fibrosis after 5/6 nephrectomy, which was accompanied by reduced expression levels of epithelial‐mesenchymal transition (EMT) markers, inflammatory mediators, and oxidative stress markers in the remnant kidneys. Treatment of cultured proximal tubular cells, HK‐2 cells, with FXa protein led to increased expression levels of EMT markers, inflammatory mediators, and oxidative stress markers, which were abolished by pretreatment with edoxaban. Treatment of HK‐2 cells with edoxaban attenuated FXa‐stimulated phosphorylation levels of extracellular signal‐regulated kinase (ERK) and NF‐κB. Our findings indicate that edoxaban can improve renal injury after subtotal nephrectomy by reducing EMT and inflammatory response, suggesting that FXa inhibition could be a novel therapeutic target for CKD patients with atrial fibrillation.
| INTRODUCTION
According to the increase of aging population, the prevalence of chronic kidney disease (CKD) continues to increase, and CKD patients are estimated to be 8%-16% of the adult population all over the world (Jha et al., 2013). Renal function is declined according to aging even in healthy subjects without causative diseases, such as type 2 diabetes and chronic glomerulonephritis (Imai et al., 2008). Recent evidence shows that CKD patients commonly have blood coagulation disorders, and according to the progression of the CKD stage, blood coagulation factors are increased, thereby leading to further CKD progression (Huang et al., 2017;Wang et al., 1997). However, little is known about the contribution of coagulation factors to the development of CKD.
Activated factor X (FXa) is a key regulator of both intrinsic-and extrinsic-coagulating cascades. FXa and its receptor protease-activated receptor (PAR) 2 signaling play an important role in various disorders including inflammation, fibrosis, and atherosclerosis (Grandaliano et al., 2003;Hara et al., 2015;Shinagawa et al., 2007;Zuo et al., 2015). FXa inhibitors, such as direct oral anticoagulants (DOACs), are widely used for the prevention of stroke and other thrombotic complications of non-valvular atrial fibrillation. Recently, Horinouchi et al. (2018) reported that oral administration of edoxaban, which is a FXa inhibitor, ameliorated renal fibrosis in a mouse model of unilateral urinary obstruction. Here, we investigated the therapeutic effects of edoxaban on CKD using a mouse model of 5/6 nephrectomy. Furthermore, we investigated the effects of FXa on fibrosis, inflammatory response, and oxidative stress in cultured proximal tubular epithelial cells.
| Animal and surgical procedure of subtotal nephrectomy
Male wild-type (WT) mice in a background of C57BL/6J at the age of 8 weeks were subjected to subtotal (5/6) nephrectomy operation, as previously described (Hayakawa et al., 2015;Ohashi et al., 2007). Briefly, the upper and lower poles of the left kidney (two-thirds of the left kidney) were resected. After 1 week, the remaining right kidney was removed through a right paramedian incision after ligation of the right renal artery, vein, and ureter. Seven days after ablation, WT mice were fed normal diets containing edoxaban (25 mg/kg/day) or vehicle for 7 weeks. Eight weeks after ablation, WT mice were sacrificed for analysis. Before the surgical procedure, anesthetization (medetomidine, midazolam, and butorphanol at doses of 0.15, 2.0, and 2.5 mg/ kg, respectively) was administered intraperitoneally. The adequacy of anesthesia was confirmed by the lack of a toepinch withdrawal response during the surgical procedure. Study protocols were approved by the Institutional Animal Care and Use Committee at Nagoya University.
| Laboratory methods
At 8 weeks from operation, mice were sacrificed for analysis. Collected blood and urine samples were used for analysis. Plasma concentrations of urea nitrogen (UN) and creatinine (Cr) and urine concentrations of Cr were measured in a commercial laboratory (SRL). Urinary albumin concentration was measured by a murine albumin enzyme-linked immunosorbent assay (ELISA) kit (Exocell). Urinary albumin excretion was evaluated as albumin/gram of urinary Cr. Plasma FXa concentration was measured by a murine FXa ELISA kit (MyBioSource).
| Histological analyses
Tissue samples were fixed by 4% paraformaldehyde and embedded in paraffin. Serial tissue sections (5 μm) of the
News and Noteworthy
Edoxaban treatment ameliorated renal function after subtotal nephrectomy with accompanying increases in EMT markers, inflammatory mediators, and oxidative stress markers. Consistently, treatment of HK-2 cells with FXa increased EMT markers, inflammatory mediators, and oxidative stress markers, which were abolished by edoxaban treatment. Finally, FXa treatment increased phosphorylation levels of ERK and NF-κB, which were abolished by edoxaban treatment through PAR2-dependent mechanisms. Thus, edoxaban can improve renal injury by reducing EMT and inflammatory response. kidney were stained with hematoxylin and eosin (H-E) and Masson's trichrome (Sigma). In immunohistochemistry, the kidney tissues were stained with antibodies for fibrinogen (Thermo Fisher Scientific). Intraglomerular cell number, glomerular size, fibrinogen positive area, and fibrosis area were measured by using an ImageJ analysis system (Ohashi et al., 2007).
| Cell culture
HK-2 cells were purchased from the American Type Culture Collection. The cells were cultured in medium consisting of DMEM/F12 (Gibco) supplemented with 10% FBS at intervals of 3-4 days to continuously passaged. HK-2 cells were placed in DMEM/F12 medium for 16 h and cultured in the presence or absence of edoxaban (100 μmol/L) for 1 h, followed by stimulation with FXa (100 nmol/L) or vehicle for 24 h. In some experiments, HK-2 cells were treated with vehicle or U0126 (20 μmol/L), which is an ERK inhibitor, for 1 h, followed by stimulation with FXa (100 nmol/L) or vehicle for 24 h. Knockdown of PAR2 was achieved by siRNA transduction at 25 nM with lipofectamine RNAiMAX (Invitrogen) 24 h before experiments. Lipofectamine RNAiMAX and siRNAs were dissolved in Opti-MEM (Gibco). ON-TARGETplus siRNA SMART pools targeting PAR2 were purchased from Horizon Discovery. Control cultures were transfected with unrelated scrambled siRNA (ON-TARGET plus Control Non-Targeting Pool, Horizon Discovery).
PCR methods
Gene expression levels were quantified by the real-time PCR method. Total RNA was extracted with RNeasy-Mini Kit (Qiagen) according to the manufacturer's protocol (Ogura et al., 2012). Extracted RNA was reversetranscribed by using the ReverTra Ace (Toyobo). PCR procedure was performed with a Bio-Rad real-time PCR detection system using THUNDERBIRD SYBR qPCR Mix as a double-standard DNA-specific dye. Primers are listed in Table 1. All results were normalized to 36B4.
| Western blot analysis
Tissue or cell samples were homogenized in lysis buffer (Cell Signaling Technology) containing 1mM PMSF (Sigma). Equal amounts of plasma were separated by denaturing SDS-PAGE and transferred onto PVDF membranes. Membranes were incubated with the primary antibodies, followed by incubation with the HRP-conjugated secondary antibodies. ECL prime system (GE Healthcare) was used for the detection of the protein signal. The protein expression levels were determined by measurement of the band intensities using ImageJ software (National Institute of Health, USA) (Schneider et al., 2012).
| Statistical analysis
Data are presented as mean ± S.E.M. The differences between two groups for variables with normal distributions were evaluated by unpaired Student's t-test. Differences between three or more groups were evaluated using one-way analysis of variance (ANOVA), with a post hoc Tukey's test. The differences between groups for variables with non-normal distribution were analyzed by the Steel-Dwass test (for three or more groups). Data distributions were evaluated by Shapiro-Wilk test. p < 0.05 denoted the presence of a statistically significant difference. All statistical analyses were performed using JMP Pro 15 software (SAS).
| Edoxaban attenuates renal damage and intraglomerular fibrin deposition after subtotal nephrectomy
To examine the effects of edoxaban on renal injury, wildtype (WT) mice at 8 weeks of age were subjected to 5/6 nephrectomy and randomly assigned to two groups, normal diet or edoxaban admixture diet at day 7 after operation. After 7 weeks, mice were sacrificed for analyses after the collection of urine and blood samples. Subtotal nephrectomy significantly increased plasma FXa levels in WT mice ( Figure S1). Plasma FXa levels were remarkably reduced in edoxaban (Edo)-treated WT mice after subtotal nephrectomy compared with vehicle (veh)-treated WT mice ( Figure S1). Subtotal renal ablation significantly increased urinary albumin excretion and circulating levels of urea nitrogen (UN) and creatinine (Cr) levels in WT mice ( Figure 1a). Urinary albumin excretion and plasma UN levels were reduced in edoxaban-treated WT mice after subtotal nephrectomy compared with vehicle-treated WT mice (Figure 1a). In contrast, plasma Cr levels after renal injury did not significantly differ between edoxabantreated and vehicle-treated WT mice.
To assess glomerular damage after subtotal nephrectomy or sham operation, kidneys of WT mice were stained with Hematoxylin and Eosin (H-E). Edoxaban-treated WT mice exhibited reduced glomerular cross-sectional area and intraglomerular cell number after subtotal nephrectomy compared with vehicle-treated WT mice ( Figure 1b).
To evaluate the effect of edoxaban on intraglomerular microembolism after subtotal nephrectomy or sham operation, kidney tissues of WT mice were immunochemically stained with anti-fibrinogen antibody. Intraglomerular microembolism shown as fibrin deposition was increased after subtotal nephrectomy operation compared with sham operation, whereas edoxaban treatment significantly reduced intraglomerular microembolism after subtotal nephrectomy compared with vehicle treatment (Figure 1c).
| Edoxaban ameliorates tubulointerstitial fibrosis by reducing EMT, inflammatory response, and oxidative stress
To evaluate interstitial fibrosis after subtotal nephrectomy or sham operation, the kidneys of mice were stained with Masson trichrome. Subtotal nephrectomy increased renal fibrosis area of WT mice, and edoxaban treatment reduced fibrosis area in the injured kidneys of WT mice after subtotal renal ablation compared with vehicle treatment ( Figure 2a). Consistently, edoxaban-treated WT mice had reduced expression levels of fibrosis markers including collagen I, collagen III, and transforming growth factor (TGF) β1, in the remnant kidney after subtotal nephrectomy compared with vehicle-treated WT mice ( Figure 2b). Edoxaban-treated WT mice also had reduced expression levels of EMT markers including αsmooth muscle actin (SMA), N-cadherin, and vimentin in the remnant kidney after subtotal nephrectomy compared with vehicle-treated WT mice ( Figure 2c). Furthermore, edoxaban-treated WT mice exhibited reduced expression levels of inflammatory mediators including tumor necrosis factor (TNF) α and monocyte chemoattractant protein (MCP) 1, and oxidative stress markers including gp91 phox , p47 phox , and p67 phox in the remnant kidney after subtotal nephrectomy compared with vehicle-treated WT mice (Figure 2d,e).
| Edoxaban attenuates FXAinduced increase in EMT, inflammatory response, and oxidative stress in cultured renal tubular epithelial cells
To dissect the precise mechanism by which edoxaban reduced EMT, inflammatory response, and oxidative stress in the remnant kidney after subtotal nephrectomy, human renal proximal tubular epithelial cell line, HK-2 T A B L E 1 Primers used for quantitative RT-PCR Mouse 36B4: forward 5'-GCTCCAAGCAGATGCAGCA-3' reverse 5'-CCGGATGTGAGGCAGCAG-3' Collagen I: forward 5'-GTCCCAACCCCCAAAGAC-3' reverse 5'-CAGCTTCTGAGTTTGGTGATA-3' cell were pretrated with edoxaban or vehicle, followed by FXa treatment. Pretreatment of HK-2 cells with edoxaban significantly reduced FXa-stimulated expression of EMT markers including α-SMA, N-cadherin, and vimentin ( Figure 3a). Furthermore, pretreatment of HK-2 cells with edoxaban significantly attenuated FXa-stimulated expression of inflammatory mediators, such as interleukin (IL) 6, TNFα, and MCP1 and oxidative stress markers including gp91 phox , p22 phox , and p47 phox (Figure 3b,c). These results indicate that edoxaban abolished FXa-stimulated increase in EMT, inflammatory response, and oxidative stress in proximal renal tubular cells. Because EMT and inflammatory response in proximal tubular epithelial cells is mediated by ERK and NF-κB signaling, respectively (Hu et al., 2020;Pollack et al., 2007), we assessed the effects of FXa on phosphorylation of ERK and NF-κB in HK-2 cells. Treatment of HK-2 cells with FXa time dependently increased the phosphorylation levels of ERK and NF-κB (Figure 4a). Pretreatment of HK-2 cells with edoxaban abolished FXa-stimulated increase in phosphorylation levels of ERK and NF-κB (Figure 4b). Consistently, subtotal nephrectomy increased renal phosphorylation levels of ERK and NF-κB in WT mice, and edoxaban treatment reduced phosphorylation levels of ERK and NF-κB in the remnant kidney of WT mice after subtotal renal ablation compared with vehicle treatment (Figure 4c). In addition, pretreatment of HK-2 cells with the ERK inhibitor, U0126 significantly attenuated FXastimulated expression of EMT markers including α-SMA, N-cadherin, and vimentin ( Figure S2).
Finally, to investigate the contribution of PAR2 to FXastimulated phosphorylation of ERK and NF-κB, HK-2 cells were transfected with siRNA targeting PAR2 or unrelated siRNA. Transduction of HK-2 cells with siRNA targeting PAR2 reduced mRNA expression of PAR2 by 71.4 ± 0.7%. Treatment of HK-2 cells with siRNA targeting PAR2 diminished FXa-stimulated phosphorylation of ERK and NF-κB (Figure 4d). These results suggest that FXa induces ERK and NF-κB activation through the PAR2-dependent pathway.
| DISCUSSION
In the present study, our major findings are following: (1) Oral administration of edoxaban ameliorated albuminuria, glomerular hypertrophy, tubulointerstitial fibrosis after subtotal nephrectomy with accompanying increases in expression of EMT markers, inflammatory cytokines, and oxidative stress markers, (2) treatment of cultured human proximal tubular cells, HK-2 cells, with recombinant FXa protein led to increased expression levels of EMT markers, inflammatory mediators, and oxidative stress markers, which were abolished by edoxaban treatment.
(3) Treatment of HK-2 cells with FXa protein resulted in increased phosphorylation levels of ERK and NF-κB, which were abolished by edoxaban treatment through PAR2-dependent mechanisms.
Renal fibrosis, characterized by tubulointerstitial fibrosis, causes tubular atrophy and glomerulosclerosis, thereby leading to disruption of normal kidney function. Residual fibroblasts are mesenchymal cells and function to maintain the structure of organs. On the other hand, myofibroblasts appear dedifferentiated from residual fibroblasts, podocytes, and renal tubular epithelial cells and contribute to pathological fibrosis during the process of CKD progression (Loeffler & Wolf, 2015;Zeisberg & Neilson, 2010). The dedifferentiation from tubular epithelial cells to myofibroblasts, so-called epithelial-mesenchymal transition (EMT), has been reported to associate with CKD progression in many animal CKD models (Holian et al., 2008;Lan, 2003;Shimizu et al., 2006;Zeisberg et al., 2008). TGF-β is the most representative inducer of EMT in many organs (Pardali et al., 2017;Xu et al., 2009). TGF-β promotes EMT of renal tubular epithelial cells and podocytes, thereby leading to renal fibrosis during CKD progression (Li et al., 2008;Liu, 2004). It has been reported that thrombin induces EMT and collagen production in retinal pigment epithelial cells (Bastiaans et al., 2013). On the other hand, the contribution of coagulation factors to renal fibrosis via EMT has not been elucidated. In the present study, we exhibited for the first time that FXa promotes EMT in HK-2 cells. Our data suggest that edoxaban suppresses FXainduced EMT in proximal tubular cells, thereby leading to reduction of renal interstitial fibrosis.
ERK signaling is reported to mediate EMT in several pathological conditions including diabetic nephropathy (Zhang et al., 2019). In the present study, we found that F I G U R E 2 Edoxaban reduces renal tubulointerstitial fibrosis, epithelial-mesenchymal transition (EMT), inflammatory response, and oxidative stress after 5/6 nephrectomy. (a) Histological evaluation of interstitial fibrosis. Upper panels show representative photos of the kidneys from vehicle-treated mice (Veh) and edoxaban-treated mice (Edo) after subtotal nephrectomy or sham operation as determined by Masson trichrome staining. The lower panel shows quantitative analysis of fibrosis area as measured by ImageJ. N = 5 in each group. Scale bars show 100 μm. (a) mRNA levels of fibrosis-associated factors such as collagen I, collagen III, and transforming growth factor (TGF) β1 in the kidney from vehicle-treated mice (Veh) and edoxaban-treated mice (Edo) after subtotal nephrectomy or sham operation. N = 5 in each group. (c) mRNA levels of EMT markers such as αsmooth muscle actin (SMA), N-cadherin, and vimentin in the kidney from normal vehicle-treated (Veh) and edoxaban-treated mice (Edo) after subtotal nephrectomy or sham operation. N = 5 in each group. (d) mRNA levels of proinflammatory mediators such as TNFα and MCP1 in the kidney from vehicle-treated mice (Veh) and edoxaban-treated mice (Edo) after subtotal nephrectomy or sham operation. N = 5 in each group. (e) mRNA levels of oxidative stress markers such as gp91 phox , p47 phox , and p67 phox in the kidney from vehicle-treated mice (Veh) and edoxaban-treated mice (Edo) after subtotal nephrectomy or sham operation. N = 5 in each group. One-way ANOVA with Tukey's multiple comparisons test (b, c (N-cadherin and vimentin), d and e) and Steel-Dwass test (a and c (α SMA)) were used to produce the p values. N.S, not significant Vimentin mRNA levels N-Cadherin mRNA levels enhanced ERK phosphorylation was observed in the remnant kidney at 8 weeks after 5/6 nephrectomy operation. A previous report showed that ERK phosphorylation is upregulated even at 12 weeks after 5/6 nephrectomy (Ding et al., 2015). Thus, it is plausible that subtotal nephrectomy induces continuous activation of ERK in the remnant kidney, thereby leading to the development of renal fibrosis, although further investigation will be required to elucidate the precise role of ERK in CKD progression.
CKD patients show higher circulating levels of inflammatory mediators, including TNFα and IL6 (Kitada et al., 2011;Upadhyay et al., 2011;Vinuesa et al., 2006). Proinflammatory cytokines, such as TNFα, promote renal injury and dysfunction, which are accompanied by increased expression of reactive oxygen species (ROS) (Bertani et al., 1989;Navarro-Gonzalez & Mora-Fernandez, 2008). In the present study, the expression of inflammatory mediators and oxidative stress markers is increased in the remnant kidney after subtotal nephrectomy compared with the kidney in sham-operated mice, whereas edoxaban administration reduced inflammatory response and oxidative stress in the remnant kidney after nephrectomy. Furthermore, FXa treatment increased the expression of inflammatory mediators and oxidative stress markers in cultured renal tubular cells. Edoxaban abolished the increased expression of inflammatory mediators and oxidative stress markers stimulated by FXa treatment. Therefore, it is likely that edoxaban can attenuate kidney injury after subtotal nephrectomy partly through the reduction of inflammatory response and oxidative stress.
Coagulation factors promote tissue injury through PARs (PAR1-4). FXa cleaves the N terminus of PAR2 and augments EMT and inflammation through MAPK or NF-κB signaling, respectively (Rothmeier & Ruf, 2012). PAR2 is abundantly expressed in the kidney and contributes to the progression of many kidney diseases, such as diabetic nephropathy, crescentic glomerulonephritis, and IgA nephropathy (Grandaliano et al., 2003;Madhusudhan et al., 2016;Moussa et al., 2007;Oe et al., 2016). In the present study, we found that FXa increased phosphorylation levels of NF-κB and ERK in cultured renal tubular epithelial cells through the PAR2dependent mechanism. Edoxaban blocked FXa-induced increase in EMT, inflammatory response, and oxidative stress as well as NF-κB and ERK activation in renal tubular epithelial cells. Of note, edoxaban diminished phosphorylation of ERK and NF-κB in the remnant kidney after subtotal renal ablation. Thus, it is plausible that edoxaban can mitigate renal injury by antagonizing the ability of FXa to enhance renal interstitial fibrosis, inflammation, and oxidative stress through PAR2/ERK or PAR2/ NF-κB signaling ( Figure S3).
In conclusion, we have shown that FXa inhibitor can ameliorate renal function in subtotal nephrectomy model, which is accompanied with reducing tubulointerstitial fibrosis and expression of inflammatory mediators and oxidative stress markers. Thus, FXa inhibition could be a novel therapeutic target for CKD patients with atrial fibrillation.
| Translational significance and limitation
We used only one type of CKD model in this study. Although the mouse model of 5/6 nephrectomy is widely used for the study of CKD, this model in C57BL/6 mice does not develop progressive renal dysfunction (Leelahavanichkul et al., 2010). Our data indicate that edoxaban reduced urinary albumin excretion in C57BL/6 mice after 5/6 nephrectomy without affecting plasma Cr levels, suggesting that this may be due to the use of mild renal dysfunction model. Further investigation will be required to clarify the effect of edoxaban on severer renal dysfunction using models of advanced CKD including glomerulonephritis and lupus nephritis.
In the present study, we found the renoprotective effects of edoxaban using a model of mild renal dysfunction. Because direct oral anti-coagulant drugs including edoxaban are prohibited to use among patients with end-stage renal disease, our findings indicate that edoxaban could be beneficial for the early-stage CKD in patients with atrial fibrillation. the experiments, acquired the data, analyzed the data, provided expertise related to the experiments, and wrote the manuscript. HO, NO (N. Otaka), HK, TT, YO, KT, and MT (M. Tatsumi) conducted the experiments, acquired the data, and analyzed the data. MT (M. Takefuji) and TM designed the research study and provided expertise related to the experiments. NO (N. Ouchi) designed the research study, analyzed the data, provided expertise related to the experiments, and wrote the manuscript. | 4,813.2 | 2022-03-01T00:00:00.000 | [
"Medicine",
"Biology"
] |
On Cross-Layer Design of AMC Based on Rate Compatible Punctured Turbo Codes
,
Introduction
The success of current standard such as 3GPP HSPA and IEEE 802.11/16 in terms of high data rates provision and quality of service (QoS) requirements satisfaction is principally owed to Adaptive Modulation and Coding (AMC), hybrid automatic repeat-request (HARQ) and fast scheduling [1,2].The AMC realization uses different constellation orders and coding rates according to the signal strength [3].By this way, when instantaneous channel conditions are proper, link adaptation offers high data rates at the physical layer.The proper usage of each constellation order and coding rate, i.e., mode is specified by the SNR regions in which each separate mode is active.
Enhancement of AMC performance can be achieved by using different channel coding techniques.Particularly, in case of turbo-coding implementation, an AMC scheme can achieve the highest spectral efficiency even if low SNR regions are met [4].The original rate of a turbo code could be 1/3; nevertheless by using puncturing techniques greater code rates can be used for each modulated symbol.Incorporating also rate compatibility in punctured turbo codes, by which all of the code sym-bols of a high rate punctured code are used by the lower rate codes, an enhanced spectral efficiency is reached.This gain is actually provided by Rate-Compatible Punctured Turbo (RCPT) codes [5].RCPT codes have been employed for HARQ implementations due to the fact that no received information is discarded [6].Such ARQ schemes are well-known as Incremental Redundancy (IR) HARQ schemes that improves the channel use efficiency since parity bits for error correction are transmitted only if this is required [7].
The aforementioned description is basically a crosslayer combination of AMC at the physical layer and HARQ at the data link layer for QoS provisioning in wireless communication networks [8,9].It has been shown that such a cross-layer design outperforms in terms of spectral efficiency, the case of AMC use only at the physical layer or the combination of typical ARQ with a single modulation and coding scheme [8].Moreover, it has been proved that IR HARQ based on convolutional codes has much improvement in spectral efficiency than that with type-I HARQ [9].To this direction and since a lot of research work has been done on turbo codes as well as turbo coding and decoding is applied to all known standards of wireless communications [2,10], we extend this study by employing the aforementioned crosslayer design (CLD) that combines AMC and HARQ based on RCPT codes.
The rest of this paper is organized as follows.Section 2 presents the system model and its components in details.In Section 3, the cross-layer design of the system is presented with its assumptions and constraints.In Section 4, system performance is evaluated for both turbo and convolutonal rate compatible codes and LDPC as well.Besides, the complexity performance is evaluated for each coded system.Finally, Section 5 provides the concluding remarks and gives some directions for further investigation in this topic.
System Model
The model of the adopted cross-layer design system is illustrated in Figure 1.It shows the layer structure of the system as well as the implementation details of the AMC scheme (i.e.physical layer).In the following text, we first describe concisely the functionality of each layer and in sequel we go into details for each of layers' components.
Turbo Encoding and Decoding
First, confirm that you have the correct template for your paper size.This template has been tailored for output on the A4 paper size.If you are using US letter-sized paper, please close this file and download the file for "MSW US ltr format".Turbo coding and decoding achieves performances on error probability near to Shannon limit [11].In its main form, turbo coding is a channel coding type that combines two simple convolutional codes in parallel linked by an interleaver (i.e.Parallel Concatenated Convolutional Codes-PCCC) [12].It had been studied that recursive systematic convolutional codes (RSC) are superior to nonrecursive counterparts for concatenated implementations [13].The codewords of such schemes consist of one information bit followed by two parity check bits which both parallel encoders produce.Thus, the rate code of a PCCC scheme with two RSC constituent codes is 1/ 3 c R .On the other side, the decoding process of concatenated codes is performed by a suboptimum decoding scheme that uses a posteriori probability (APP) algorithms instead of using Viterbi algorithms.Such a scheme is constructed by "soft-in/soft-out" decoders that exchange bit-by-bit or symbol-by-symbol APPs as soft information that depends on the bit or symbol decoding technique [14].The input soft-information represents the log-likelihoods of encoder input bits and code bits.This is actually the input of the Soft-Input Soft-Output (SISO) "Maximum A Posteriori" (MAP) module presented in [15].The output soft-information of this module is updated versions of input based on the information of the constituent RSC of the turbo encoder.
More specific, turbo decoding based on a PCCC scheme is constructed by two SISO modules that linked with a deinterleaver (Figure 2).In addition to that, iterative decoding is accomplished in order to improve the decoding performance.Henceforth, a feedback loop between the two constituent SISO decoders is established that actually presents the turbo decoding principle [11].This feedback loop appears an interleaver that gives interleaved inputs to the first parallel decoder required for the first iteration and so on.Multiple iterations between these two decoders exchanging soft information give near to Shannon limit results.Turbo codes and iterative turbo decoders has been extensively studied for implementation purposes in current standards like 3GPP HSPDA (High Speed Data packet Access) [16].
RCPT Codes
In general, a RCPT encoder consists of a turbo encoder as described above followed by a puncturing block with puncturing matrix P. The puncturing matrix P is known as the puncturing rule or pattern and indicates the coded bits that should be punctured [17].Puncturing can be applied both to information or/and parity bits.However, the way of puncturing affects the coding scheme performance and the coded-modulation scheme in general [18].Assuming only the impact of puncturing on turbo coding scheme, one can realize that without puncturing systematic bits, the code performance decrease is reasonable.In addition to that, by puncturing periodically the parity bits produced by two RSC codes, a better performance of the coding scheme can be achieved.
The rate compatibility offered by a RCPT code has been considered as the enabling technique for incremental redundancy (IR) HARQ schemes [6].IR HARQ based protocols are major components of HSDPA offering rate matching capabilities [19].During the rate matching process, the transmitter sends only supplemental coded bits indicated by the aforementioned puncturing rule.A representative example of IR HARQ scheme for HSDPA with turbo encoder as mother code is presented in [20].The RCPT encoder in particular is constructed by a turbo mother code with a rate code resulted by RSC encoders.The puncturing matrix indicates the puncturing period and actually the bits being punctured during the HARQ scheme operation [6,18].Therefore, the resultant family of rate codes is: An example of RCPT encoder dedicated to ARQ mechanisms with M = 3 and P = 3 is illustrated in Figure 3 which is constructed from two constituent RSC encoders with rate 1/2 and offers a family of RCPT code rates
RCPT-ARQ Protocol
By puncturing the bits that will be transmitted in the current and future transmission attempts, the HARQ scheme (i.e.RCPT-ARQ) brakes the packet unit with size into blocks of bits with size .The number of transmitted bits of the RCPT-ARQ protocol at the transmission attempt can be expressed as ) 1, denotes the rates produced by the RCPT encoder.Going into further details, we assume a single stop-and-wait ARQ strategy of RCPT-ARQ protocol (i.e.hybrid ARQ) described by the following step-by-step functionality: C Depending on previous channel condition the adaptive scheme operates on mode n.
The L-long packet size is encoded by the turbo mother code.The coded packet is stored at the transmitter and is broken into blocks with size of {1 / } . Bits selection is performed for each transmission attempt according to the puncturing rule.
Constituent blocks' transmission with size i L is initialized according to the puncturing matrix.
At the receiver, iterative decoding is performed for each separate transmission block i L .If decoding is not successful after the number of maximum transmissions max t N is reached, then a NAK is sent to the transmitter and the adaptive scheme updates to the corresponding mode according to the channel condition.
Otherwise, an ACK is sent to the transmitter and the adaptive scheme continues to the current mode n.
Cross-layer Design
The cross-layer system structure described above is relied on the following assumptions: Channel SNR estimation is perfect and in consequence the channel state information (CSI) that is available at the receiver as well, although the impact of errors in SNR estimation on adaptive modulation is negligible [21].In our implementation, the channel estimator is implemented using the Error Vector Magnitude (EVM) algorithm for AWGN channel [22].
Feedback channel dedicated to mode selection process is error free and without latency.The mode selection is performed in a packet-by-packet basis i.e. the AMC scheme is updated after max t N transmission attempts.Alternative update policies with e.g.updates for every transmission attempt (i.e.block-by-block basis), will be left for further investigation [10].
System updates are based on received SNR denoted as that is actually the estimated channel SNR at the receiver.It is assumed that the received SNR values per packet is described statistically as i.i.d random variables with a Rayleigh probability density function (pdf): where { } E is the average received SNR.
Cross-Layer Design of AMC and HARQ
A cross-layer design approach that combines the AMC at the physical layer with a hybrid ARQ at the data link layer could follow the procedure presented both in [9] and [10].Applying this method, the following constraints must be imposed in order to keep a particular QoS level at the application layer: Constraint1 (C1): The maximum allowable number of transmissions per information packet is .C1 is calculated by dividing the maximum allowable delay at the application layer and the round trip time required for each transmission at the physical layer.For example, assuming the QoS concept of 3GPP, the audio and video media streams for MPEG-4 video payload allows a maximum delay value equal to 400 ms [23].In addition to that the round trip delay between the terminal and the Node B for retransmissions in case of HSDPA could be approximated less than 100 ms [23].Thereafter, in such a context, the should be 4. On the other hand, C2 is related to the bit-error rate (BER) at the physical layer and the packet size at the data link layer.Hence, if the BER imposed by the QoS requirements at the application layer is equal to and the information packet size is L = 1000 then the should be It is obvious that the aforementioned cross-layer design (CLD) dictates the code rates that will be used for each transmission at the data link layer and therefore specifies the AMC switching thresholds at the physical layer.Moreover, the proposed CLD scheme will be affected by constituent encoders (i.e.RSC encoders) of turbo code as well the puncturing rules [6,18].However, in current investigation, we present the results derived using one of the optimal RCPT code and puncturing rule presented in [6], and we will present the RCPT codes and puncturing impact on our CLD in our future work.
AMC Schemes
The design of AMC schemes is the process by which the switching thresholds are specified.The switching thresholds of an AMC scheme at the physical layer are specified by a given target BER ( ) [3,4].The switching thresholds are boundary points of the total SNR range denoted as Afterwards, each mode is selected in accordance to the switching thresholds derived from the .
arg et t BER However, in a combined system in which the unit of interest is the packet at the data link layer, the AMC design follows the value.More specific, in order to satisfy the aforementioned constraints of the proposed combination, the switching thresholds should be derived from the following inequality: where is the packet error probability (i.e.packet error rate) after transmissions at the data link layer.In the following paragraphs, we derive the boundary points for each modulation and coding scheme (MCS).
The packet error probability can be expressed in rela-tion to BER by the following equation only if each demodulated and decoded bit inside the packet has the same BER and bit-errors are uncorrelated [9,10].On the other hand, known closed form expressions for the PER1 and BER is not available in the literature and closed-form expressions for the BER of turbocoded modulations in AWGN channel is not available either [8].All the same, one can use the union bound for turbo-coded modulation system using the bounding technique introduced in [24].However, this technique is applied for 16QAM system and indeed needs more investigation in case of turbo-coded AMC schemes with multiple modulation modes.Thereafter and since further investigation on union bounds of turbo-coded modulation is not the aim of our current work, we take BER and PER values through simulations.Finally, the simulated PER values are compared with those derived from fitting the curves and those derived from Equation (5).
Figure 4 shows the PER values versus received SNR of each mode in coding step with , where i Rc 1, 2, 3 i number of transmissions.We use the 1/2 QPSK, 3/4 QPSK, 1/2 16QAM and 3/4 16QAM modulations with RCPT code rates 1, 1/2 and 1/3 for each transmission respectively.The packet size is length and the puncturing follows the optimal rules according to [6] (Table 1).The constituent RSC encoders of PCCC turbo codec is the optimal encoder B proposed by [6] g have octal representations and respectively.The number of iterations is 8.The figures depict the simulated PER, the fitting curves and the values derived from Equation ( 5).(15) octal (13) octal In order to have a more clear view on RCPT performance combining with AMC, we should compare it with the other types of rate compatible codes.To this end, we implement also the aforementioned CLD first using RCPC (Rate Compatible Convolutional Code) and second using RC-LDPC (Rate Compatible Low-density Parity-check codes).We use the same rates for both two RC codes.Specifically, the RCPC is a convolutional encoder with rate 1/2, generator polynomial (171, 133) and constraint length 7 [9].For LDPC, we employ the same codes as in [25] with rate 1/2 (1008,504) and a variable node degree equal to 3. The corresponding performance of these modulation and coding schemes (MCS) is depicted separately for each code in
System Performance
In case of a general type-II HARQ that uses punctured codes, the probability of unsuccessful reception after t N nsmissions represents the event of decoding failure with code i Rc ter i transmissions [10].In case of limited transmissions, the packet error probability of this using AMC mode n N under channel states is given by [26] 1,..., By using ( 6) over for each retransmission and for each mode the packet error rates after transmissions are resulted.The denotes the region boundaries for each MCS and obtained as follows The is reached using the corresponding decoder when the imposed transmission attempts is reached either.Assuming , the derived switching thresholds are listed in Table 2. Table 2 includes also the parameters of MCSs for convolutional and LDPC codes.We next evaluate the system performance in terms of spectral efficiency when the AMC scheme is combined with type II HARQ (i.e.IR HARQ).In each transmission attempt, the number of transmitted bits is specified according to RCPT code rates , where s T S is the symbol rate.Afterwards, the spectral efficiency gives the bit rate in bits per symbol that can be transmitted per unit bandwidth and is given by In (7), where is the input information packet size and L L is the average of transmitted symbols in order to transmit an information packet.The average of transmitted symbols for each mode is given by n For cross-layer designed AMC schemes with n = 1,..,N modes, the average spectral efficiency needs to be calculated in order to evaluate system performance.By averaging the n L values in the range of for over all ( ) (1) ( ,..., ) modes, the average number of transmitted symbols in order to transmit an information packet is rate compatible punctured codes are employed under the constraints of the previous described cross-layer design.The parameters of each MCS are those listed in Table 2 considering a channel with Rayleigh fading phenomena as described above.
In Figure 6, we make contrast of the average spectral efficiency derived for each rate compatible punctured code.We illustrate the values of third transmission (i.e.N t = 3).Figure 6 shows the performance merit of RCPT against RCPC.This corroborates the benefit of turbo scheme against convolutional one in terms of communication performance as it is well known.Indeed, this performance benefit is more evident in low regions of average SNR than in high regions.Moreover, it is obvious that RC-LDPC achieves performance close to RCPT code.This is a useful outcome considering these two families of codes since LDPC codes are used in several standards and especially in space communications.The fact that turbo and LDPC codes show identical performance has also concluded both in [27] and [28].[27] has focused on performance in terms of PER values at the physical layer both in AWGN and multipath Rayleigh fading channel.[28] has proposed the PEG (Progressive Edge-Growth) construction method for LDPC codes and has concluded that turbo coding is identical of LDPC in terms of bit-level performance.To this direction, we evaluate the system performance under the aforementioned cross-layer design and we have also concluded in the same result.
Comparison Complexity
However, the comparison between different codes should not be considered only in terms of performance related to communication efficiency.It should be also studied in terms of complexity even when the achieved system performance is identical between different codes (e.g.turbo and LDPC).Most of code complexity issues are related to computational complexity measuring the additional operations required by each code.Another important aspect of code complexity relies on architectural issues introduced by code design.[29] studies the complexity of decoding algorithms that is measured in terms of computational operations such as multiplications, divisions and additions.In Table 3 is listed the number of operations (i.e.additions, divisions, etc.) needed for each decoding procedure using the max-log MAP (Maximum A Posteriori) algorithm and the Viterbi algorithm in case of turbo and convolutional decoder respectively.These are actually the decoding algorithms that we have implemented in the RCPT and RCPC decoding procedure.In Table 3, M is the constraint length used by each encoder at the transmitter side.
Figure 7 shows the complexity of each decoding procedure (i.e.turbo and convolutional) in terms of number of operations vs. the number of iterations and code constraint length respectively.It is obvious from this figure that the decoding complexity in case of convolutional scheme is noticeably less than turbo case.In our case, the convolutional decoding procedure uses Viterbi decoder with constraint length equal to seven.On the other hand, turbo decoding uses max-log MAP with iterations equal to eight.The declension of turbo decoding complexity is close to two times the complexity of convolutional one since convolutional decoding scheme exhibits 1200 number of operations while turbo one exhibits approximately 2400 number of operations.On the other hand, performance comparisons between turbo and LDPC codes in terms of decoding complexity have shown that when both codes achieve an identical performance then the decoding complexity remains approximately the same.For instance, [28] have claimed that 80 iterations using the belief propagation algorithm produces the same decoding complexity as a turbo code does with 12 iterations using the BCJR decoding algorithm.[27] has studied the performance comparison between turbo and LDPC codes in more details considering computational complexity.The authors have measured the computational complexity in terms of number of operations per iteration per information bit that they could be additions or comparisons.Table 4 shows the computational complexity per information bit of the sub-optimum decoding algorithm for code rate R = 1/3.The complexity is expressed in relation to number of iterations and it is illustrated in Figure 8.
itr N Assuming the same configuration as in [27] the turbo decoding with 8 iterations when a max-log-MAP algorithm is used exhibits approximately the same complexity in terms of number of additions with the LDPC decoding scheme that uses the BP algorithm.In our comparative study, we use the decoding schemes from [28] that consist of a turbo decoder with max-log-MAP plus 8 iterations and LDPC decoder with PEG decoding graphs plus 80 iterations.Henceforth, it could be claimed that both turbo and LDPC decoders show the same computational complexity.
Conclusions and Future Work
In this paper, we have extended the cross-layer design combining AMC with HARQ using RCPT codes.To this end, a hybrid FEC/ARQ based on RCPT codes has been assumed.In previous works, the proposed CLD was introduced with uncoded modulations, convolutional and rate-compatible convolutional coded modulations dedicated to AMC schemes.In addition to that, we have implemented a CLD approach using puncturing techniques for rate compatibility purposes.The system performance has been evaluated for type-II hybrid ARQ mechanism.Moreover, we have illustrated comparative results of system performance of other rate compatible codes as convolutional and LDPC as well.In order to have a more comprehensive view of coding and decoding schemes we also discuss the computational complexity of each code separately, in terms of the required number of operations either in each iteration attempt or for each memory length.However, since turbo coding and indeed punctured turbo codes are able to accomplish better performance with different RSC encoders and puncturing rules namely optimal encoding and puncturing [26], a future work should be the performance evaluation of AMC and HARQ combination implementing different encoders and puncturing rules using RCPT-ARQ.
Figure 1 .
Figure 1.Cross-layer system model combining AMC and HARQ based on RCPT codes.
): The probability of unsuccessful reception after transmissions is no greater than .max t N Pr loss
2 .
In addition to that, when mode is used, each transmitted symbol carry n As in[9], we assume a Nyquist pulse shaping filter with bandwidth s B T
Figure 5 Figure 5 .
Figure 5 depicts the average spectral efficiency of the combination of AMC and type-II HARQ relied on constraints and .In this figure, it is shown the performance of AMC at physical layer when 3 t N 0.01 loss PER
Figure 6 .
Figure 6.Comparison of RC codes in terms of average spectral efficiency under the constraints of CLD design. | 5,238.8 | 2010-03-31T00:00:00.000 | [
"Computer Science",
"Engineering"
] |
Proton tautomerism for strong polarization switching
Ferroelectrics based on proton tautomerism are promising in low-field and above-room-temperature operations. Here seven organic ferroelectric crystals are examined to search for efficient switching of strong spontaneous polarization on proton tautomerism. Solution-grown crystals exhibit strong pinning of ferroelectric domain walls, but excellent switching performance is awakened by depinning domain walls under thermal annealing and/or repetitive bipolar pulses with a high voltage. Compared with ferroelectric polymers such as polyvinylidefluoride, the optimized polarizations are comparable or stronger in magnitude whereas the coercive fields are two orders of magnitude weaker. The polarization of croconic acid, in particular, breaks its own record for organic systems in increasing from 21 to 30 μC cm−2 and now exceeds those of some commercial ferroelectric materials such as SrBi2Ta2O9 and BaTiO3. Optimization reduces the discrepancy of the spontaneous polarization with the results of the first-principles calculations to less than 15%. The cooperative roles of proton transfer and π-bond switching are discussed by employing the point-charge model and hydrogen-bond geometry.
F erroelectrics are electrically polar substances in which the direction of spontaneous polarization is reversibly switchable under the influence of an external electric field. Owing to the strong polarization and high Curie point, ferroelectric oxides have gained prominence for many practical applications in information technology, such as non-volatile memory, capacitors, sensors, actuators, ultrasonic devices, and nonlinear optics applications [1][2][3][4] .
Organic systems are free of both toxic and rare elements, well suited to highly productive printing (low-temperature solution process) into sheet devices, and are expected to be advantageous for emerging applications with cheap, disposal, flexible, wearable and/or implantable characteristics [5][6][7][8] . Organic ferroelectrics have found some piezoelectric applications such as sensors and acoustic devices, although their available compounds have been limited to polyvinylidefluoride (PVDF, (CH 2 CF 2 ) n ) variants. In contrast, low-voltage-memory operation has long been a challenging issue for polymer ferroelectrics because of the high switching field exceeding several hundred kilovolts per centimetre. Recently, some small organic molecules have appeared as future potential alternatives; molecules in which proton transfer within the hydrogen bonds can invert the crystal polarity under much lower switching fields typically ranging from one to several tens of kilovolts per centimetre 9,10 . As a consequence, few-volt switching has been achieved even on printed micrometre-thick single-crystal films of 2-methylbenzimidazole (MBI) 11 .
Among ferroelectric small molecules, croconic acid (CRCA) 12 has materialized the strongest polarization through a cooperative proton tautomerism mechanism [13][14][15] (called prototropy, abbreviated herein as PTM), which relocates a proton through the hydrogen bond and simultaneously interchanges the locations of a single bond and adjacent double bond (Fig. 1). This discovery was followed by the development of PTM ferroelectrics using the b-diketone enol O ¼ C-C ¼ C-OH moieties (that is, 2-phenylmalondialdehyde (PhMDA) and 3-hydroxyphenalenone (HPLN)) 16 , carboxylic acid O ¼ C-OH moiety (that is, cyclobutene-1,2-dicarboxylic acid (CBDC)) 16 , and heterocyclic -N ¼ C-NH-moieties (that is, MBI and 5,6-dichloro-2methylbenzimidazole (DC-MBI)) 17 . All these PTM ferroelectric crystals construct extended chains of intermolecular resonanceassisted hydrogen bonds (RAHB) 18 , which are strengthened by the interplay with the conjugated p-bond system. The findings of ferroelectricity support the concept proposed by Haddon,Stillinger,and Carter 19 in 1982 that the hydrogen-bonded PTM system would be one of the representative models for memory functional molecular devices. In addition, the large polarization with significant electronic contribution has been closely related to the observations of very strong second-order nonlinear optical effects of CRCA such as second harmonic generation and terahertz radiation due to optical rectification 20,21 .
Despite having excellent prospects, many organic ferroelectric crystals including those based on PTM have often exhibited some ambiguity in actual performance; polarization-electric field (P À E) hysteresis loops are often ill-shapen, and the muchlower-than-expected remanent polarization values depend strongly on the pristine crystals 22 . Recently, the investigation of the domain-property relationship by Kagawa et al. 23 has provided useful clues for the refinement of performance. In general, ferroelectric crystals adopt a multidomain structure in which ferroelectric domain walls (DWs) separate differently polarized sections (domains) and their sweeping motion changes the bulk polarization. Kagawa et al. 23 visualized the domain structures of a solution-grown crystal of acid-base alternating supramolecules by piezoresponse force microscopy (PFM) and demonstrated that imperfect switching arises from charged DWs being less mobile against the external field than neutral DWs. Herein, 'neutral' or 'charged' denotes the presence or absence of bound charges on DWs and depends on whether the relative angle between the domain-wall plane and the polarization vector P in the domains is parallel or antiparallel 23 . To avoid a huge depolarization field on the DWs, dense bound charges must be almost completely compensated for by mobile charges and/or immobile charged defects. It should be noted that these pinning sites (that is, charged DWs) are likely to be embedded simultaneously with the crystal growth at room temperature but can be eliminated by reconstructing the domain structures once after heating beyond the Curie point. Removal of charged DWs and compensating trapped charges turns out to be one strategy for efficient switching of the polarization in above-room-temperature ferroelectric supramolecular crystals.
Similar DW pinning is expected for solution-grown PTM ferroelectric crystals, because their Curie point is above room temperature or even beyond the thermal stability limit of the solid state. Sotome et al. 21 imaged changes in the domain structures of CRCA crystals under various electric fields by measuring the emission of terahertz radiation and found that charged DWs were always almost fixed in space and prevent bulk switching of the polarization. Previous P À E hysteresis measurements of CRCA 12 exploited both thermal annealing and repetitive switching for depinning DWs, although poor reproducibility of the best polarization (21 mC cm À 2 ) suggested that these procedures are still insufficiently optimized. Similarly, the P À E hysteresis loops of a CBDC crystal exhibited a polarization switching that was small in comparison with the theoretical results (2.9 versus 6.6 mC cm À 2 along the c-direction) 16 .
Here we revise the P À E hysteresis properties of solutiongrown PTM ferroelectric crystals by finding effective optimization procedures. The optimized remanent polarizations (P r ) were evaluated in comparison with the results of firstprinciples calculations. We considered the cooperative roles of proton transfer and p-bond switching using the iondisplacive (point-charge) picture and hydrogen-bond geometry. Seven PTM ferroelectric compounds including a new compound allow for a systematic understanding of polari- zation in search of design principles for high-performance switching.
Results
Materials and structural assessment. Our previous studies have discovered six PTM ferroelectrics: CRCA, PhMDA, HPLN, CBDC, MBI and DC-MBI 12,16,17 . All these crystal structures comprise extended one-or two-dimensional networks through intermolecular hydrogen bonds; O-H?O bonds for the former four compounds and N-H?N bonds for the latter two. Their crystal symmetry and three-dimensional molecular packing are revisited in Supplementary Table 1 and Supplementary Figs 1-5 along with the crystal polarities. As shown schematically in Fig. 1, the polarity of the hydrogen-bonded networks is switchable through cooperative proton transfer and concomitant interchange of the double-and single-bond locations (the interconverting bonds and atoms involved are specified in red). In addition, the crystallographic requirement for ferroelectricity is a hidden pseudo-symmetry, which would survive as a paraelectric configuration.
Here, we also report a N-H?N bonded compound as the seventh PTM ferroelectric: 3-anilinoacrolein anil (ALAA) 24 , which was encountered in the Cambridge Structural Database (CSD) (Refcode: ANPHPR). Because this entry does not include the atomic coordinates for the hydrogen atom, we have re-examined the crystal structure at room temperature (T ¼ 295 K). The crystal structure belongs to the orthorhombic system with the polar space group Iba2 (#45) in agreement with the report. The crystal polarity is parallel to the crystal c-axis, which is along the direction that the hydrogen bonds construct an infinite chain of twists and turns. Without the N-H proton and p-bond alternation, the molecule can restore the pseudo-twofold rotation symmetry and occupy the C 2 -site so as to constitute the centric (hypothetical) paraelectric structure: the space group Ibca (#73). Although the hydrogen-bonded chain acquires a dipole switchable with the proton location, its polarization should mostly arise from the switchable p-bond dipole because all the protons travel in a direction normal to the polar axis. The orthogonality of the chain dipole and the proton's path is similar to the case of the HPLN crystal ( Fig. 1 and Supplementary Fig. 6).
The possibility of 3-hydroxydibenzo[a,c]tropone (DBT, Supplementary Fig. 7a) exhibiting ferroelectricity was previously conjectured solely on the basis of earlier structural analysis of the polar crystal symmetry (monoclinic; space group Cc) 25 . In this investigation, DBT was synthesized according to the literature 26 and crystallized from solution. The P À E hysteresis experiments unexpectedly revealed a signature of antiferroelectricity along the O-H?O bonded zigzag chain parallel to the crystal c-direction ( Supplementary Figs 7 and 8). Through careful structural reassessments using a synchrotron X-ray source, antiferroelectric behaviour was attributed to the symmetry reduction from a C-centred to primitive lattice having antiparallel chain polarity arrangements. For details of the structural and electric characterizations, Supplementary Figs 7 and 8 and Supplementary Discussion.
Theoretical polarization versus previous experiments.
According to the current dielectrics theory of solids, the evaluation of macroscopic polarization through the Berry phase formalism demands precise knowledge of the electronic structures in the crystal form 27,28 . Therefore, for all seven ferroelectrics, first-principles electronic structure calculations were performed to evaluate the spontaneous polarization P cal . For the target ferroelectric structure (l ¼ 1), the atomic coordinates of all the non-hydrogen atoms were obtained from X-ray diffraction data. For the hydrogen atoms, the core locations were computationally relaxed so as to minimize the total energy, because the X-ray diffraction merely locates the corresponding electron density maxima with underestimated X-H (X ¼ C, O or N) distances and a large ambiguity. 29,30 . The reference hypothetical paraelectric structure (l ¼ 0) was constructed from this ferroelectric structure by adding inversion symmetry. Supplementary Table 2 summarizes the changes in the space-and point-group symmetries between the ferroelectric and hypothetical paraelectric structures. The electronic structures and corresponding spontaneous polarizations were calculated for different degrees of polar distortion l between the centric (hypothetical paraelectric, l ¼ 0) and fully polar (ferroelectric, l ¼ 1) configurations.
For each compound, the validity of the simulations was confirmed by the smooth l-dependence of the polarization ( Fig. 2) as well as the polarization vectors lying along the symmetrically allowed direction. The P cal values of the CRCA and CBDC crystals are close to the corresponding theoretical predictions: (P a , P b , P c ) ¼ (0, 0, 26.0) mC cm À 2 for CRCA 12 and (P a , P b , P c *) ¼ (12.7, 0, À 6.6) mC cm À 2 for CBDC 22 . These values change slightly depending on the details of the computational conditions such as the structural parameters and the exchange-correlation functional. For reference, Picozzi et al. 31 reported polarizations of 24-32 mC cm À 2 for CRCA using various exchange-correlation functionals.
In Table 1, we compare the theoretical polarizations P cal with the P À E hysteresis data P exp . For the PhMDA, MBI and DC-MBI crystals, the remanent polarizations listed are those of the previous report 16,17 and agree well with the corresponding theoretical polarizations. Since all these crystals were commonly grown from a high-temperature vapour phase, the thermal annealing effect is likely to have maximized the performance by eliminating pinned DWs. In fact, PFM images of MBI crystals revealed only neutral 180°and 90°DWs rather than charged DWs 17 .
However, for all the solution-processed as-grown single crystals, the P cal values (Table 1) are much larger than previously reported values 16 : CRCA (P c ¼ 21 mC cm À 2 ), HPLN (3.0 mC cm À 2 , normal to the 10 1 ð Þ plane), and CBDC ((P a , P c ) ¼ (0.9, 2.9) mC cm À 2 ). The crystals were grown from slow evaporation of solution at close to room temperature: CRCA from 1 N hydrochloric acid, a-form HPLN from ethanol, and CBDC from water. These observations, indicative of the strong DW pinning, necessitated us to excite DW motion through a preconditioning treatment such as thermal annealing and/or repetitive switching.
Optimization of solution-grown ferroelectrics. All the previous measurements and optimizations of polarization switching of CRCA, HPLN, and CBDC crystals were conducted in air or inert gas. The present re-examinations have reinforced the maximum amplitude (E max ) of the bipolar electric field, which was hitherto set to less than 33 kV cm À 1 for a triangular waveform so as to avoid an electric discharge between electrodes on the crystal surface. Satisfactory optimization has been accomplished for crystals immersed in a silicone oil with field amplitudes increased up to 33-100 kV cm À 1 using rectangular-pulse voltages (Fig. 3).
Earlier work on a CRCA crystal increased P r up to 21 mC cm À 2 by applying 600 cycles of a 1 Hz triangular-wave voltage of E max ¼ 33 kV cm À 1 followed with thermal annealing at 400 K (ref. 12). Here, we applied bipolar rectangular-pulse voltages of a stronger field amplitude (pulse field amplitude E puls ¼ 33-55 kV cm À 1 ) to four specimens of different initial P r ranging from 2 to 13 mC cm À 2 . Gradual expansion of the hysteresis loops with an increasing number of pulses corresponds to the so-called wake up process 32 . After a few ten thousand pulses, the P r reproducibly reached a maximum as high as 28-32 mC cm À 2 ( Supplementary Fig. 9). This optimum P r is in excellent agreement with our theoretical evaluation (29.4 mC cm À 2 ) and also with the recent studies of Picozzi et al. 31 Moreover, the switchable polarization breaks its own record for organic ferroelectrics by increasing from 21 to about 30 mC cm À 2 and has just exceeded the performance of some commercial ferroelectric materials such as BaTiO 3 and SrBi 2 Ta 2 O 9 (20-26 mC cm À 2 ) 1,33 . For the crystal specimen exemplified in Fig. 3a, the optimized P À E loops are improved in their rectangularity and frequency independence of P r in comparison with the corresponding earlier data. It should be noted that such strong polarization can be fully switched with a low coercive field (34 kV cm À 1 ) even at a frequency as high as 1 kHz.
Polarizations of the a-form HPLN crystal were improved with both thermal heating and repetitive switching ( Supplementary Fig. 10). Although the calculated polarization vector is parallel to the a-axis, only well-developed crystal surfaces available for experiments were the 10 1 ð Þ and 101 ð Þ planes ( Supplementary Fig. 3c). Through thermal annealing, we noticed that the ferroelectric phase is thermally robust at least up to T ¼ 425 K, at which temperature the P À E curves are quasi-rectangular with P r ¼ 5.0 mC cm À 2 and E c B10 kV cm À 1 at frequencies of |P exp | (P x , P y , P z ) exp |P cal | (P x , P y , P z ) cal |P ion | (P x , P y , P z ) ion 1. 30-1,000 Hz ( Supplementary Fig. 10c). The switchable polarization decreased slightly after cooling and additionally with time at room temperature, but it could be optimized to P r ¼ 4.5 mC cm À 2 again by continuously applying bipolar rectangular-pulse voltages of a strong field (33 kV cm À 1 ) at room temperature (Fig. 3d).
Considering the inclination angle between P cal (B||a) and the applied field E, the observed component of P exp along E corresponds to a total amplitude |P exp | of B5.6 mC cm À 2 . Substantial improvements have also been achieved in switching the CBDC crystal ( Supplementary Fig. 11). The crystal polarity is uniaxial and lies within the crystallographic ac plane by symmetry. Both the a-and c*-direction polarizations were maximized by applying stronger-field cycles: bipolar triangularwaveform voltages (E max ¼ 60 kV cm À 1 ) and rectangularpulse voltages (E puls ¼ 45 kV cm À 1 ), respectively (Fig. 3b,c). The optimized polarization P exp exhibited P a ¼ 8.6 mC cm À 2 and P c * ¼ 10.0 mC cm À 2 , which did not depend on frequencies up to 30 Hz. This P exp is comparable to the calculated P cal in both magnitude (13.2 versus 15.2 mC cm À 2 ) and direction, as depicted by the open bold arrows in Supplementary Fig. 11c. It should be noted that these polarization vectors are almost parallel to the hydrogen-bonded chain running along the crystal 201 ½ direction (solid arrow).
The ALAA crystal, which was newly grown from ethanol solution, also exhibited ferroelectricity in the P À E hysteresis experiments with an E||c configuration (Fig. 3e), in agreement with the structural assessment above. The remanent polarization was similarly optimized to a modest value (3.6 mC cm À 2 ) after applying 3 Â 10 4 cycles of triangular-waveform voltages (E max ¼ 60 kV cm À 1 ). The ferroelectric state is stable at least up to the melting point at 388 K, below which the signature of the phase transition was absent until room temperature according to a thermal analysis using a high-sensitivity differential scanning calorimeter ( Supplementary Fig. 12b).
For the solution-processed as-grown PTM ferroelectric crystals, the switchable polarization has been 'woken up' by repetitive switching with an increased amplitude of the bipolar field rather than by improving the crystal quality. Hence, most of the pinning sites are not permanently clamping impurities or defects. Rather, they are charged DWs with compensating charges trapped nearby and can be gradually moved away under influence of a strong electric field. Note that the remanent polarizations become independent of the applied field frequency after optimization in all hysteresis loops (Fig. 3a-e). After the 'wake-up' process, polarization fatigue started with a steep increase in the coercive field E c beyondB10 5 cycles for CRCA, CBDC, and ALAA and beyond B10 6 cycles for HPLN. These 'wake-up' and fatigue behaviours are quite analogous to those of domain depinning and fatigue observed in some hard ferroelectric oxides 32,34-38 . In the latter ferroelectrics, the observed changes in the remanent polarization, coercive field, frequency dependence, and loop curvature have been well described by a model incorporating the variable interaction strength between the switchable dipole and fixed dipole as well as the depolarization field 35 . Similar arguments might be applied to the switching mechanism for ferroelectric PTM.
All the solution-processed as-grown PTM ferroelectric crystals herein has been finally optimized with significant improvement of their spontaneous polarization. In comparison with the corresponding earlier data, the improved performance is also evident by the optimized P À E curvatures themselves (Fig. 3). Good rectangularity of the loop and frequency independence of P r suggested the almost complete removal of charged DWs and then successful disclosure of the genuine materials' properties.
Comprehensive comparison of polarizations. Figure 4a plots the best experimental performance |P exp | of the seven PTM compounds as a function of the theoretical polarization amplitude |P cal | along with that of anthranilic acid (ATA) form I 39 , another hydrogen-bonded above-room-temperature ferroelectric. Because the applied field E is inclined with respect to the predicted P cal for the HPLN and MBI crystals, each measured P exp was corrected to the total amplitude |P exp | considering this inclination angle (arrows in the figure). Note that the polarizations vary widely from 3.6 to 30 mC cm À 2 . After the optimization and correction, all the experimental data fall near the linear line |P exp | ¼ |P cal |, and the largest discrepancy, found for the ALAA and CBDC crystals, is only 14-15%.
During the switching with PTM, the protons travel B0.6-1.0 Å within the hydrogen bond. This classical picture suggests significant ionic polarization. The line graph in Fig. 4b actually reveals a trend that shows the total polarization |P cal | (solid diamonds) increasing with proton density (solid squares). We first estimate this electrostatic contribution from protons displacing among molecular (anionic) cores. In the point-charge model, the ionic polarization P ion is expressed as where e is the electron charge, u i is the relative displacement of the static charges Z i |e|, O is the unit cell volume, and Z i is taken as þ 1 for protons and À 1 or À 2 for molecular cores. The negative point charges were placed at the centre of gravity of the p-conjugated cores: the pentagon C 5 O 5 2 À for CRCA, b-diketone enol C 3 O 2 À for PhMDA and HPLN, ethylenedicarboxylate C 4 O 4 2 À for CBDC, imidazole ring C 3 N 2 À for MBI and DC-MBI, and bridging C 3 N 2 À unit for ALAA. The positive point charges were placed on protons at the energetically optimized locations calculated above. The u i of each point charge is the displacement from the hypothetical paraelectric structure constructed by imposing pseudo-symmetry elements on the ferroelectric structure. Table 1 lists the calculated P ion and the magnitude of the local dipole moment |l| of the crystallographically independent protons. The dipole moment l i was calculated from the proton displacements u i by In the histogram in Fig. 4b, each total polarization P cal was divided into P ion (red bar) and a remainder contribution (blue bar). Each molecule of CRCA or CBDC accommodates two protons in the compact molecular size, and the resulting large proton density amplifies P ion . However, P ion is still less than half of P cal in amplitude, indicating the addition of some larger contributions. This is true also for the MBI, DC-MBI, and PhMDA crystals. It should be noted that the HPLN and ALAA crystals reveal a nearly zero P ion due to the orthogonality of the proton's path and the chain dipole. Their large polarizations then manifest from the significant contribution of the remainder mechanisms. The most important contributions to P other than P ion should come from the sections that experience the most dramatic redistribution of charge during the polarity reversal. The corresponding origin is nothing but switchable p-bond dipoles, the heart of PTM, which interchanges the locations of a single bond and adjacent double bond. Note that P exp and P cal of ALAA are smaller than those of HPLN despite very similar molecular size and switchable p-bond fragment. The reduced polarization could be explained by the significantly inclined orientation of switchable p-bond dipole from the bulk polarization vector (Supplementary Fig. S6).
Comparison of switching field. For hydrogen-bonded ferroelectrics such as KDP (KH 2 PO 4 ) and its isomorphs, the physical properties are closely related to the local hydrogen-bond geometry, which critically affects the potential barrier height for proton hopping between two equilibrium positions 1,40,41 . Likewise, a series of ferroelectric supramolecules of anilic acids exhibited positive relationships in which stretching the hydrogenbonded length enhanced both the phase-transition temperature and polarization performance 42 . Although the hydrogen bond lengths vary across the seven PTM ferroelectrics, they cannot be related to the thermal stability because of the lost paraelectric state 43 . As noted above, the bulk polarization strongly depends on whether the proton motion is nearly parallel or normal to the crystal polar axis. In turn, we found structural effects on the switching field. The P À E hysteresis loops provide the coercive field E c , which is determined by the field at the P ¼ 0 intercept and is accompanied by the peak (displacement) current in the This observation suggests that proton hopping would be a rate-limiting process during the switching, which occurs at an E c two orders of magnitude lower than those of polymers such as PVDF (around several hundred kilovolts per centimetre).
Discussion
Our success in efficient optimization of ferroelectric switching has reduced the discrepancy of the polarization with the results of the first-principles calculations to less than 15%. Thus, one of the important outcomes of this work is the hallmark on the practicality of the calculations, which will be applied satisfactorily for the prediction of experimentally unknown performance of similar organic systems from the available precise crystal structure. It should be noted that evaluation of the ferroelectricity as well as the theoretical polarizations require careful diffraction studies. The unexpected finding of antiferroelectric DBT has demonstrated that routine structural assessment is not always straightforward. The other attractive findings are the strong optimized polarization and its low-field switching. Compared with ferroelectric polymers such as PVDF, the polarizations of the CRCA and CBDC crystals are stronger and those of the other compounds are comparable in magnitude. The CRCA crystal even breaks its own record for organic systems and also beats some commercial ferroelectric materials. The PTM ferroelectrics studied herein exhibited a stable ferroelectric state up to temperatures well above room temperature even until the stability limit of the crystals themselves, such as the melting, decomposition, or sublimation temperatures. Two issues are relevant to this thermal stability. The first is the necessity of the depinning process for the solution-processed asgrown crystals. This is because the solution process, at far below the Curie point, spontaneously grows polarized crystals of a multidomain structure (that is, twinning) and embeds charged DWs frozen therein. The second is the excellent consistency between the experimental and simulated polarizations. One of the reasons for this is the deep potential minima in the fully polarized state, which minimized the ambiguity in the proton locations once accurate positions of the non-hydrogen atoms were experimentally determined.
The structure-property relationship together with the pointcharge-model analysis yielded some molecular and crystal design strategies for higher-performance PTM ferroelectrics. The first issue concerns the direction of each contributing dipole moment. The total polarization P cal was described by the accumulation of the ionic contribution P ion from the proton displacement and the larger remaining contributions mainly from switchable p-bond dipoles. These cumulative contributions to the polarization are just converse to the subtractive nature in intramolecular hydrogen-bonded PTM such as 9-hydroxyphenalenone: the p-bond dipole changes in the opposite direction to that of the relocating protons, and some cancellation yields tiny switchable molecular dipoles (0.4 DB1.3 Â 10 À 30 C m) 44 compared with the local dipole moment around the hydrogen atom 45 . In contrast, the extended chains of intermolecular hydrogen bonds are ideal for enlarging the polarization, especially when both the local proton motion and the change in the p-bond dipoles are nearly aligned parallel to the polar crystal axis. This requirement is well satisfied in the CRCA, CBDC, and DC-MBI crystals.
The second issue is the density of dipoles. As noted above, the total polarization actually increases with increasing proton density. Similarly, one can easily envisage that the corresponding strategy for switchable p-bond dipoles is to increase the volume fraction of PTM fragments, which are specified in red in Fig. 1. As a typical case, the dibasic acid CRCA and CBDC are similar in terms of the ionic contribution P ion and the proton density, but the remainder contribution in CRCA is huge and about three times as large as that of CBDC. The reason for this can be understood qualitatively in that the only fragment irrelevant to PTM is one C ¼ O unit in CRCA and much smaller than the cyclobutene C 4 H 4 unit in CBDC. To summarize, the recordbreaking performance of CRCA can be attributed to the ideal arrangement of dense dipoles; that is, the effective addition of ionic and p-bond polarizations, the highest spatial density of protons, and spatially dense PTM fragments extending across nearly the whole C 5 O 5 core.
There are a number of useful tools established herein for further development of organic ferroelectric materials: the molecular design principles, structural assessment, theoretical simulation, and optimization procedures considering DW dynamics. Our results indicate that the optimization of molecular, crystal, electronic, and/or domain structures can bring more insight into the goals of higher performance, new functionalities, and useful applications of organic systems.
Methods
Sample preparation. ALAA purchased from Alfa Aesar was recrystallized twice from slow evaporation of ethanol solution at 5°C to afford new ferroelectric crystals of elongated orange plates. DBT was synthesized according to the literature 26 , purified by vacuum sublimation in the temperature gradient, and crystallized from cold methanol. CRCA from Tokyo Chemical Industry was recrystallized three times from 1 N hydrochloric acid solution evaporated slowly under a stream of argon gas, and thereby collected as yellow plates. HPLN from Acros Organics was purified by a few repetitions of vacuum sublimation in the temperature gradient. CBDC purchased from Wako Pure Chemical Industries was recrystallized three times from acetonitrile. Orange plates of HPLN (a-form) and colourless plates of CBDC were grown under slow evaporation of the ethanol and aqueous solutions, respectively. For other compounds, we adopted almost the same purification and crystallization procedures as those in previous work.
Crystallographic studies. The X-ray diffraction data collection for the ALAA crystal at room temperature and the assignment of the crystallographic axes of the bulk single crystals were completed using a four-circle diffractometer equipped with a hybrid pixel detector (Rigaku AFC10 with PILATUS200K; graphitemonochromated MoKa radiation). The intensity data were analysed with the CrystalStructure crystallographic software packages (Molecular Structure Corp. and Rigaku Corp.). The final refinements were done with anisotropic atomic displacement parameters for the non-hydrogen atoms and with a fixed C-H bond length of 0.95 Å for the hydrogen atoms. (7) crystals. Solid curves are guides for the eye. Each bond distance is obtained by averaging over crystallographically independent sites. NATURE COMMUNICATIONS | DOI: 10.1038/ncomms14426 ARTICLE Electric measurements. All the electric measurements employed single crystals with painted gold paste as the electrodes. The P À E hysteresis curves were measured by the virtual ground method 46 using a ferroelectrics-evaluation system (Toyo Corporation, FCE-1), which consists of a current/charge-voltage converter (Toyo Corporation Model 6252), arbitrary waveform generator (Biomation 2414B), analogue-to-digital converter (WaveBook 516), and voltage amplifier (NF Corporation, HVA4321). All the crystals were immersed in silicone oil to avoid electric discharge with a maximum electric field exceeding 30 kV cm À 1 . The high-temperature measurements and/or annealing were also conducted by heating the sample in the oil bath.
First-principles calculations. First-principles computational code QMAS 47 based on the projector augmented-wave method 48 and the plane-wave basis set was employed for calculations of spontaneous polarization P cal through the Berry phase formalism 49,50 . To describe the electronic exchange-correlation energy, the Perdew-Burke-Ernzerhof (PBE) version of the generalized gradient approximation (GGA) was used 51 . The target ferroelectric structures (degrees of polar distortion l ¼ 1) were constructed from the atomic coordinates of all the non-hydrogen atoms determined by the previous X-ray diffraction studies at room temperature. The locations of the hydrogen atoms were computationally relaxed so as to minimize the total energy. Refcodes of the corresponding CIF files, which we previously deposited in the CSD and used for calculations herein, are GUMMUW02 (CRCA), PROLON01 (PhMDA), TAPZIT01 (HPLN), CBUDCX01 (CBDC), KOWYEA (MBI) and REZBOP (DC-MBI).
Data availability. X-ray crystallographic data have been deposited with the Cambridge Crystallographic Data Centre (CCDC) under deposition numbers CCDC-1498921-1498923 and can be obtained free of charge from the Centre via its website (www.ccdc.cam.ac.uk/getstructures). All other data supporting the finding of this study are available within the article and its Supplementary Information. | 7,375.2 | 2017-02-16T00:00:00.000 | [
"Physics",
"Materials Science"
] |
Evaluation of synthetic data generation for intelligent climate control in greenhouses
We are witnessing the digitalization era, where artificial intelligence (AI)/machine learning (ML) models are mandatory to transform this data deluge into actionable information. However, these models require large, high-quality datasets to predict high reliability/accuracy. Even with the maturity of Internet of Things (IoT) systems, there are still numerous scenarios where there is not enough quantity and quality of data to successfully develop AI/ML-based applications that can meet market expectations. One such scenario is precision agriculture, where operational data generation is costly and unreliable due to the extreme and remote conditions of numerous crops. In this paper, we investigated the generation of synthetic data as a method to improve predictions of AI/ML models in precision agriculture. We used generative adversarial networks (GANs) to generate synthetic temperature data for a greenhouse located in Murcia (Spain). The results reveal that the use of synthetic data significantly improves the accuracy of the AI/ML models targeted compared to using only ground truth data.
Introduction
Modern technologies provide sustainable and feasible solutions to many real-world problems.One area where these technologies have provided solutions in recent years is agriculture.Precision agriculture applies innovative technologies to the agricultural world to reduce costs, increase profit and achieve sustainability [1].A comprehensive review of the state of the art use of artificial intelligence (AI) in smart greenhouses is provided by [2].This review focused on the optimization of crop yields, reduction of water consumption, fertilizers, diseases, pests, and the search for improved agricultural sustainability.Therefore, the status of various AI technologies in smart greenhouses is reviewed by discussing the extent to which technologies have been successfully applied in an agricultural context and the options for optimizing their usability.
Among the challenges facing precision agriculture is the adaptation of processes to climate change [3].To monitor crop status to face sudden weather changes that occur mainly B Juan Morales-García<EMAIL_ADDRESS>author information available on the last page of the article in semi-arid climates, farmers use technologies such as the Internet of Things (IoT) to monitor their plots and/or greenhouses [4,5].The data generated by these systems also feed into decision support systems to perform intelligent and automatic actions on the plots.Several leading examples for these include climate control in greenhouses [6] or frost prevention in a fruit orchard through smart irrigation [7].
Although decision support systems have numerous advantages and can make decisions in anticipation of future climatic conditions, they have the disadvantage of needing to create local models to achieve high accuracy in predicting climate variables [8,9].This disadvantage translates into the need to have historical data on the location of the plot to train and create an accurate model according to the farmer's needs.This would mean installing the IoT system to collect data but not accurately using the prediction system until there is sufficient historical data to create the prediction model.In [10], the authors review four bio-inspired intelligent algorithms used for agricultural applications, such as ecological, swarm intelligence-based, ecology-based, and multi-objective-based algorithms.Some observed that no universal algorithm could perform multiple functions on farms; therefore, different algorithms were designed according to the specific functions to be performed.
Despite being in the era of Big Data, there is still a lack of quality data to address local problems such as the one mentioned above [11].Recently, AI techniques have emerged that can generate artificial data of equal or higher quality than the original data, thus solving the problem of the amount of data needed to train local models [12].Among these techniques, generative adversarial networks (GANs) -deep artificial neural networks capable of generating artificial data - [13] have obtained interesting results in different applications, including image processing [14], speech recognition [15] and other [16].
Within the field of precision agriculture, GANs have recently been applied to image processing tasks such as image augmentation [17,18] and other tasks within computer vision [19].However, to the best of our knowledge, synthetic data generation has not been applied to time series data generation in precision agriculture for climate control.In this study, we propose and evaluate synthetic data generation strategies to increase the accuracy of forecasting models for greenhouse climate control.
Greenhouses are agricultural structures that must be tightly controlled to avoid extreme weather conditions to achieve high crop yields [20].Therefore, farmers are increasingly installing greenhouses controlled by IoT systems to monitor their crops in real time.However, using these data to generate a greenhouse climate model that allows intelligent and automatic control to reduce resources used while increasing crop production is challenging.Therefore, to develop this predictive model, the historical data set to train this model is crucial.These data are not available for the specific location where the greenhouse is installed until the IoT system starts operating.To solve the data problem, this study proposes the creation of synthetic greenhouse data using GAN techniques, to design a prediction system for climatic variables, specifically focusing on temperature, as it is one of the most influential monitored variables [21].The findings of this study include: • Creation of synthetic datasets using GANs techniques considering different time granularities.• Study of the best prediction technique using neural networks to predict the temperature of a greenhouse, considering various granularities.• Analysis and comparison of the different models created with both synthetic and original data, as well as with the fusion of both types of data.
The remainder of the paper is organized as follows.Section 2 summarizes state-of-the-art related studies regarding synthetic data generation in a time series.Section 3 describes the proposed GAN technique for creating synthetic time series data, as well as the techniques used for evaluating such synthetic data, including the description of the data and evaluation metrics used for the assessment.Section 4 shows the results, analysis and discussion.Section 5 highlights the conclusions and directions for future works
Related works
Data collection and capture is one of the mayor features of an open and well-served society.Innovative technologies allow us to capture, analyze and merge data from a variety of sources.However, data are not always accessible, because of privacy or because there is no local data collection system for a problem [22].In this situation, new AI technologies provide tools and techniques capable of creating synthetic data.Synthetic data is a simulation of ground truth data that allows us to have a greater amount of information, to obtain more robust and accurate techniques [23].When creating synthetic data, it is important to consider the type of data to be created.The creation of synthetic image data is useful and is widely used for health problems [24] or disease detection in crops [25].However, the need for larger data sets is not exclusive to the world of image processing.Furthermore, in all contexts that require data for ad-hoc training, they also require large datasets, whether regarding IoT (where time series data predominates) or open contexts (where tabular data predominates).In [26], the authors review the role of IoT devices in smart greenhouses and precision agriculture, where variables such as the cost of agricultural production, environmental conservation, ecological degradation and sustainability have been analyzed.It shows how the economic benefits of using IoT applications in smart greenhouses have long-term benefits in commercial agriculture.
Focusing on the generation of synthetic data for time series data, synthetic data generation methods based on long-short term memory (LSTM) techniques are widely used.In [27], using LSTM, a method for completing synthetic well logs from existing log data was established.This method allowed, at no additional cost, synthetic logs to be generated from input log datasets, considering variation trend and context information.Furthermore, combining standard LSTM with a cascade system was proposed, demonstrating that this method gives better results than traditional neural network methods, and the cascade system improved the use of a stand-alone LSTM network, providing an accurate and cost-effective way to generate synthetic well logs.
Another of the most widely used techniques for synthetic data generation in recent years is GANs [28].The use of GANs in time series has been widely used to detect anomalies, both in univariate [29][30][31] and multivariate models [32].This scheme is widely used when working with unsupervised learning where anomaly detection is of particular importance for class labeling.The works on synthetic generation of time series data are not focused on agriculture; they are general works where techniques are proposed and evaluated with benchmarks or work focused on other areas.Yoon et al. [13] proposed a framework for the generation of synthetic time series data, where supervised and unsupervised techniques are combined.Specifically, the authors propose an unsupervised GAN with supervised training using autoregressive models.
However, in agriculture, using time series GANs is rarely used.Some studies have used agricultural data as benchmark data [33,34], but to the best of our knowledge, there are no publications that focus on solving precision agriculture problems using GANs.In this study, the usefulness of synthetic data is investigated by assessing whether they preserve the distribution of individual attributes, the accuracy of the ML models and pairwise correlation.
Materials and methods
This section shows the datasets used and their characteristics.The synthetic data generation model was introduced before AI models were used to validate the effectiveness of the synthetic data described.Finally, different training strategies followed to achieve the objective are presented.
Dataset
The creation of synthetic data must first take a ground truth dataset from the particular domain for which synthetic data will be generated.In this case, the actual data are obtained from an operational greenhouse located in a semi-arid region of south-eastern Spain (Murcia).ground truth data is obtained from an IoT infrastructure that measures the inside temperature (ºC) of this greenhouse, which has been in continuous operation since 2018.This infrastructure sends 5 minutes of data grouped into 15-minutes, 30-minutes and 60-minutes respectively by performing the standard average.
Because the greenhouse is located in a semi-arid region, the thermal differences between summer and winter are remarkable; therefore, it has been considered that the ground truth data should be divided into winter and summer periods as well.Table 1 shows the ground truth datasets we have created for evaluation purposes.It shows the starting and ending date of the data, and the total number of values available.Datasets ending with a W indicate the end of the training data in winter and datasets ending with an S indicate the end of the training data in summer.
Synthetic data generation using GANs
For the generation of synthetic data, this study used Doppel-GANger; a GAN architecture for sequential data proposed in [35].Figure 1 shows the GAN architecture used that is based ., R S in Fig. 1).
According to authors, this allows us to better capture the temporal correlation of long series and reduce the number of passes required by the model to generate the synthetic data.Furthermore, the GAN also includes a normalization mechanism for each input time series to tackle the well-known model-collapse problem of many GAN models.Then, the discriminator, which is a multilayer perceptron (MLP) with up to five layers of 200 neurons each followed by a ReLU activation function, uses the Wasserstein loss to report the differences between the ground truth and the fake data.
Deep Learning models
To assess the impact on the accuracy of ground truth and synthetic time series, four deep learning models have been considered: (1) MLP, (2) CNN, (3) LSTM and (4) a combination of CNN and LSTM.• MultiLayer Perceptron (MLP): The multilayer perceptron is an artificial neural network made up of multiple layers that forms a directed graph through the different connections between the neurons that make up the layers.This neural network attempts to simulate the biological behavior of neurons.MLP can solve non-linearly separable problems, because each neuron, apart from the inputs, has a non-linear activation function.The MLP is based on the backpropagation method.This method attempts to adjust the weights of the network connections to minimize the prediction error between the output produced by the network and the desired output.Layers can be classified into three types: The input layer comprises the neurons that input the data; no computation occurs in these neurons.Hidden layers can be as numerous as necessary depending on the complexity of the data; these layers comprise neurons whose input comes from previous layers and whose output and settings are passed on to subsequent layers.Finally, the output layer comprises neurons whose values correspond to the number of outputs of the network.In this study, a three-layer MLP comprising input, hidden and output layers are used.The first receives the input features; the hidden layer is where the inputs are processed so that the output layer generates the output of the MLP.The hidden layer learns any complex relationship between the input and the output due to the activation functions of its neurons [36].
• Convolutional Neural Network (CNN):
Convolutional neural networks are a type of supervised learning artificial neural network that processes its layers by mimicking the visual cortex of the human eye to identify different features in the inputs.These layers perform operations that modify the data to understand its particular characteristics.The three most common layers are: convolution, activation or ReLU, and clustering.The convolutional layer applies a set of convolutional filters to the input data where each filter activates different features.The rectified linear unit holds positive values and sets negative values to zero, allowing for faster and more efficient training, also known as activation, as only activated features proceed to the next layer.The clustering layer simplifies the output by a non-linear reduction of the sampling rate, which reduces the number of parameters the network must learn.These operations are repeated in tens or hundreds of layers; each layer learns to identify different features.After learning features in various layers, the architecture of a CNN moves on to classification.The penultimate layer is fully connected and generates a K-dimensional vector.The final layer of the CNN architecture uses a classification layer to provide the final classification output.The difference between a CNN and a traditional neural network is that a CNN has shared weights and bias values, which are the same for all hidden neurons in a given layer.Although the use of convolutional neural network models is more associated with the image classification domain, they are also used in different applications and domains, such as regression, where they can be used with time series by transforming the data to adapt them to the input of the convolutional network [37].
• Long Short-Term Memory (LSTM):
The LSTM model has a recurrent neural architecture with state memory, having the advantage of allowing long-term memory, and is therefore widely used in time series.LSTM is an evolution of standard recurrent neural networks, used in machine learning problems where time is involved, because their architecture as cells and loops allows the transmission and recall of information in different steps.
LSTM comprises an architecture that allows information to be stored over long time intervals.This is because the memory cells of the network comprise several layers with loss functions (instead of one as in usual recurrent networks) of sigmoid type that allow us to bypass or add information to the main information line of the neural network, controlled by a hyperbolic tangent function.The information passes from one cell to another, first passing through a sigmoid layer, which is called the forget gate layer.It compares input and output, and returns a value between 0 and 1.If it is 1, the information is stored, if it is 0, it is disregarded.The next step comprises the second sigmoid layer and the hyperbolic tangent layer.It is used to decide which new information will be stored in the cell.The sigmoid layer called the input gate layer decides which value will be updated, and the hyperbolic tangent layer creates a vector of possible values decided by the previous one to be added to the state.The last step is a sigmoid layer that decides what the output will be, followed by a hyperbolic tangent layer that decides which values go to the network output according to the sign by which they are multiplied [38].
Preparation of datasets for training and testing
To accurately assess the impact of the synthetically generated data, five training and testing strategies are proposed to assess the performance of the ML models previously presented.The first strategy (that is, the ground truth dataset) is based only on the ground truth dataset (see Section 3.1).This dataset is divided into two datasets: (1) the training dataset, comprising all the data except the last day, and (2) the test dataset, comprising the last day of the available data.As these are time series data, it is impossible to perform a cross-validation or a validation with any other dataset than the latest values of the time series.time series require preserving the order and dependence between the data.The second strategy for training and testing (namely, Synthetic dataset) only relays on the synthetic data generated with the GAN model previously presented.The synthetic dataset is divided into two datasets: (1) the data used for training, i.e., the synthetic data generated and the data used for testing that, in this case, are obtained from the ground truth dataset and (2) the data used for testing; i.e., the last day of the time series.The evaluation data are removed, and instead, the evaluation data are taken from the ground truth dataset, so the impact of the synthetic data on a real scenario can be rigorously evaluated.
The third strategy (namely, Synthetic + Ground truth dataset) combines synthetic and ground truth data.The ground truth dataset has been extended by adding data at the beginning of the dataset from the synthetic dataset to extend the time series and thus increasing the size of the dataset for training.Likewise, the models are trained using the entire dataset described above, removing the last day, which is reserved for testing.
The fourth strategy (namely, Synthetic + Ground truth with reinforcement learning dataset) is inspired by reinforcement learning.It also uses synthetic data with ground truth data but here, the training is performed by only using synthetic data.Once the model has been trained, the model is re-trained by using ground truth data.This is because the greenhouse will be continuously operating, and thus, data will be increasingly generated.Then, it can be used to increase the performance of the models over time.Likewise, the test strategy uses the last ground truth day to evaluate accuracy.
The fifth strategy (Shuffled synthetic + Ground truth dataset) uses synthetic and ground truth datasets.This test is like the third strategy, but, the synthetic dataset is shuffled before being concatenated at the beginning of the ground truth dataset.Like previous strategies, the last day of the ground truth dataset is used for testing.This strategy is used to verify the validity of a criterion-generated time series, and it would not be valid to introduce mere random data.
Evaluation and discussion
This study considers two dimensions of the problem: (1) the use of GANs for synthetic data generation (time series data) and ( 2) the impact on the accuracy of AI models depending on whether ground truth or synthetic data are used.
Exploratory data analisys
All the hyperparameters that have been used for using the GAN model are specified, described and explained in the following list: Table 2 shows the main statistical values of the ground truth time series sampled every 15, 30 and 60 minutes during two and a half years together with the same descriptive statistics of the synthetic series over 288, 144 and 72 years.
Most are the usual statistical values.In particular, the standard error of the mean (SEM) measures how much discrepancy is likely in a sample's mean compared with the population mean.Kurtosis is the degree of peakedness of a distribution, if the value is close to 0, then a normal distribution is often assumed.Skewness is usually described as a measure of a dataset's symmetry, a value between -0.5 and 0.5, the data are fairly symmetrical.The statistics for skewness and kurtosis simply do not provide any useful information beyond that already given by the measures of location and dispersion but is another element to compare in the last column.Root-mean-square error (RMSE) is a frequently used measure of the differences between values, in our case ground truth and synthetic predicted values.
As can be observed, RMSE, calculated from the ground truth and synthetic column of each sampling rate, is a notably Fig. 2 Box plot comparing ground truth and synthetic data distributions according to sampling frequency small value for all statistical measures shown.In addition, we can check the standardised mean difference (SMD) which tests for differences in means between ground truth and synthetic time series.Normally, a value of less than 0.1 is considered a "small" difference.
Table 2 shows a notably statistical similarity between the ground truth and synthetic values, especially because so many years are artificially generated.The data to see the distribution of the time series helps identify possible numerical anomalies such as outliers that would cause similar statistical values for different distributions.That is why these conclusions must be visually corroborated by looking at the box-and-whisker diagram shown in Fig. 2, the Kernel Density Function shown in Fig. 3 and the three Q-Q plots shown in Fig. 4 that compare the ground truth (line) and synthetic To corroborate the conclusion that the generated synthetic time series will be useful to enrich the training of predictive models with tens of thousands of samples that we lack in reality, we compare on the timeline the three sets of generated series.Figure 5 shows a comparison of one week sampled every 15, 30 and 60 minutes between ground truth and synthetic data sets.
Visually, the synthetic time series is adjusted to the periodicity of each actual day.It is not perfect but significant Fig. 4 Q-Q plot comparing ground truth and synthetic data sets according to sampling frequency 123 Fig. 5 Comparison of the same week of the three sampling rates with respect to their corresponding generated time series correlations between each pair of ground truth and synthetic datasets are reported.However, they are not statistically significant when analyzing the correlation month-to-month or, year-to-year (see Table 3).A priori, this is not a problem for the intention to use the synthetic results to improve predic-
Model evaluation
Table 4 shows the models and hyperparameters used for assessment purposes.
The results of each model described in Section 3.3 using the above parameters are presented next.We have used three metrics to perform such an evaluation, the mean absolute error (MAE), the root mean squared error (RMSE) and coefficient of determination (R 2 ).These are some of the most common metrics used to measure accuracy for continuous variables.MAE and RMSE are suitable for model compar- where, y i is the real (ground truth) value of the climatological variable, ŷi is the predicted value, e 2 i is the error term and n is the number of observations.
Table 5 shows the values of the metrics for the MLP for the five train strategies described in seconds (secs.).3.4.As seen, the strategy following a reinforcement learning approach achieved the best scores in most metrics and time horizons.This is especially remarkable for the datasets with a time frequency of 15 minutes (GreenHouse-15m-W and GreenHouse-15m-S ).Furthermore, such a reinforcement approach provided more accurate MLP models than those solely relying on ground truth data.The R 2 of the former approach was 0.936 for GreenHouse-15m-S whereas the score of the latter strategy was only 0.644 given a 12-h time horizon.Similar behavior was observed for the 24-h period given the same dataset, 0.957 vs 0.835 R 2 .The strategy using a shuffled version of the synthetic time series achieved larger errors than the one combining the time series because the GAN directly generated them.Concerning the sensitivity of the results, the accuracy of the MLPs trained following the synthetic or the synthetic + ground truth policies seem to slightly decrease with the frequency increases up to 60 min.For example, the R 2 score of the synthetic dataset MLP was 0.913 and 0.886 for frequencies 15 and 30 min given the summer dataset but it dropped to 0.749 when the frequency is set to 60 min.However, this pattern is not observed in the other policies in Table 5.
Table 6 shows the results obtained from the CNN model.Here, the three strategies that incorporated synthetic data during the training stage improve results than the one solely relying on the ground truth data.The combination of synthetic and ground truth data strategies achieved the best scores for all metrics and time horizons for the GreenHouse-15m-W feed.A similar behavior was observed in GreenHouse-30m-W.However, when the frequency increased to 60 min in the winter feed (GreenHouse-60m-W), reinforcement learning or only the use of synthetic data strategies provided better results.However, the summer datasets showed, a slightly different pattern.The CNN models trained with synthetic or the reinforcement-learning strategies were more accurate for the 30-min frequency (GreenHouse-30m-S dataset), but the combination of synthetic and ground truth strategy provided the most accurate CNN model for 15-min and 60-min frequencies.This reveals that combining the synthetic with the ground truth data approach improved the training of the CNN with high time frequencies (15 min) but for lower frequencies the other two synthetic-based approaches were also suitable.In terms of sensitivity, the models following ground truth or shuffled synthetic+ground truth approaches improve results when the frequency increases from 15 min to 60 min.However, the other three approaches follow the opposite trend with a slight accuracy improvement when decreasing the frequency of the time series (e.g. the R 2 score of the MLP with Synthetic + ground truth approach moved from 0.798 to 0.869 when the frequency of the GreenHouse-60m-S decreased from 60 to 30 min.This suggests that, for the MLP model, the combination of synthetic and real data must be better considered for time series with frequencies below 30 min.Table 7 summarizes the evaluation of the LSTM model.The three synthetic-based training strategies outperformed the approach that only used ground truth data, considering most metrics, time horizons and datasets.For example, the RMSE of the LSTM trained only using ground truth data was 6.358 for the GreenHouse-15m-S dataset when considering a 24-h time horizon the same model trained with synthetic data achieved a much lower RMSE, 3.829.Furthermore, the LSTM model exhibited differences in terms of accuracy depending on the time frequency of the model, as already observed with the CNN model.Therefore, Table 7 shows that the reinforcement-learning approach allowed the LSTM model to improve its accuracy for most of the datasets with low time frequencies (GreenHouse-30m-S, GreenHouse-60m-W and GreenHouse-60m-S).Furthermore, the approach that relies solely on synthetic data to train the model generated more accurate predictions datasets with higher time frequencies (i.e., GreenHouse-15m-W and GreenHouse-15m-S) at least for the 12-h time horizon.The training strategy based on a shuffled version of the synthetic time series achieved larger RMSE and MAE values than the three versions using the original synthetic time series, as well as the LSTM model just trained only with ground truth data.Table 7 also shows that all the models trained with the four policies, including ground truth data, were sensitive to the frequency of the input time series.The R 2 score exhibited an increase in the ground truth, synthetic, synthetic + ground truth, and shuffled synthetic + ground truth policies when the frequency of the time series moved from 30 to 60 min.In contrast, a different behavior was observed for the LSTM solely trained with synthetic data, its more accurate results were obtained with the frequency of the input time series was set to 15 min.Last, Table 8 comprises the evaluation results of the CNN+LSTM model.The three training alternatives that used synthetic time series improved results, than the one that was based solely on ground truth data.Furthermore, we can see that the strategy that combined ground truth with synthetic data achieved the best results especially for the 15, or 30 min datasets.For example, the RMSE of the model for a 12-h prediction when trained was 0.932 for the GreenHouse-30m-W.This was a lower error than the one obtained by the variation trained only with ground truth data (i.e., 1.645).Furthermore, the CNN+LSTM model, trained only with synthetic data, achieves the best results for the two datasets with a 60-min frequency.Unlike the previous models, the reinforcement-learning strategy performed sligthly worse than the other alternatives.Moreover, the training using shuffled synthetic data, achieved slightly higher errors than the other four alternatives in most cases.Regarding sensitivity, CNN+LSTM variations improved scores with the 24-h time horizon than with the 12-h configuration.Furthermore, CNN+LSTM solely trained with ground truth data obtained better results for the summer than for the winter feeds considering its R 2 score (e.g., 0.928 vs 0.945 for the 60 min with 24 h as prediction horizon according to Table 8).This seasonal sensitivity was also observed in the other four policies incorporating synthetic data.In this study, there are common patterns in the results of the four evaluated models.1) The training of the forecasting algorithms leveraging the synthetic time series improved their prediction capabilities regarding the alternative of relying on ground truth data.2) Common behavior is that using a shuffled version of the synthetic data did not provided no meaningful improvement regarding the models with just ground truth data.3) The strategy combining ground truth with synthetic data provided the most robust models for 15min and 30-min frequencies, at least for the CNN and LSTM variants.For larger frequencies, the reinforcement learning strategy provided more reliable predictors.
Evaluating the strategies has also revealed a sensitivity of the models to the frequency and season of the input time series.However, how these two factors affect the accuracy of the predictors strongly varies across models and training strategies with no global sensitivity pattern.Although the MLP and CNN with ground truth data performed better in the winter season, the other alternatives with synthetic data seem to provide better results in the summer time series.However, the CNN and CNN+LSTM alternatives do not follow such seasonal trends and show slightly better results in summer than in winter, regardless of the particular training strategy used to compose the predictor.
This has important implications in operational terms as it would be necessary to consider the relevance of the season and the frequency of the time series in order to eventually select a training strategy and the predictive algorithm.For example, in the case of greenhouse settings where the summer season was the most important part of the year, the evaluation showed that a CNN or CNN+LSTM instance trained with a synthetic + ground truth policy would be the most suitable configuration.The evaluation has shown that, for example, the RMSE of the CNN+LSTM model solely trained with ground truth data was above 3.00 for all the summer feeds (Table 8) whereas the CNN+LSTM fed with synthetic and ground truth data was below 2.42 for the same summer feeds.
These findings confirm the main hypothesis of this work, the usage of coherent synthetic time series, to enlarge the training sets of a forecasting model, helps to improve their final accuracy.Furthermore, shuffled series also shows that this improvement does not occur because we added more data to the training corpus, but because of the use of a synthetic series that actually behave in a similar manner to the target one.
Conclusion and future work
Precision agriculture is moving from tele-control systems to intelligent control systems by exploiting the data generated from the IoT system for a more sustainable and efficient crop management.This transition requires substantial amounts of reliable and ready-to-use data from the deployment of the system to train ML/DL models that meet expectations.
In this context, this novel study shows the reliability and suitability of using synthetic time series to expand the training corpus of deep-learning to forecast algorithms.The goal of these algorithms is to predict the internal temperature of greenhouses to anticipate future actions to keep this internal temperature within a suitable range.Five training strategies have been defined to optimally fuse ground truth and synthetic data.
The models trained with some of these fusion strategies outperformed the alternative models solely trained with the raw measurements from the temperature sensors by considering different time frequencies, evaluation metrics and time horizons.The metrics evaluated were affected by the frequency of the target time series and the season under consideration (winter or summer).This calls for a careful procedure to select the model and the training strategy based on the period of the year under study and the characteristics in terms of frequency and data curation applied on the input sequences of data.
This work opens a novel and promising research line for studying the most suitable training strategies for combining raw and synthetic time series in the development of a smart greenhouse.Future work will focus on: 1) Developing other combinations of ground truth and synthetic data to further improve the prediction of AI/ML models; 2) Using other synthetic data generation techniques and evaluating their effectiveness; 3) Apply the transfer learning technique for time series models of synthetic data generation; 4) Generate synthetic data and AI models in their multivariate version that consider all the variables that exist in a greenhouse; 5) Apply synthetic data generation methods and AI models in contexts other than those of precision agriculture in greenhouses.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made.The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material.If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.To view a copy of this licence, visit http://creativecomm ons.org/licenses/by/4.0/.
Fig. 1
Fig. 1 Architecture of the DoppelGANger used for the synthetic data generation
Fig. 3
Fig. 3 Kernel density function comparing ground truth and synthetic data sets according to sampling frequency
Table 1
Description of ground truth dataset
•
Max sequence length: Length of time series sequences, variable length sequences are not supported, so all training and generated data will have the same length sequences.Used value is: Lenght of the time serie for one day (96, 48 or 24), deppends on the dataset.• Sample length: Time series steps to generate from each LSTM cell in DGAN, must be a divisor of max_sequence_len.Used value is: Lenght of the time serie for one day (96, 48 or 24), deppends on the dataset.Learning rate for Adam optimizer.Used value is: 0.0001.• Discriminator learning rate: Learning rate for Adam optimizer.Used value is: 0.0001.• Epochs: Number of epochs to train model.Used value is: 100000.
• Batch size: Number of examples used in batches, for both training and generation.Used value is: min(1000, length of the dataset).• Apply feature scaling: Scale each continuous variable to [0,1] or [-1,1] (based on normalization param) before training and rescale to original range during generation.Used value is: True.• Apply example scaling: Compute midpoint and halfrange (equivalent to min/max) for each time series variable and include these as additional attributes that are generated, this provides better support for time series with highly variable ranges.Used value is: False.• Use attribute discriminator: Use separate discriminator only on attributes, helps DGAN match attribute distributions.Used value is: False.• Generator learning rate:
Table 2
Comparison of ground truth and synthetic temperature time series distribution
Table 3
Average of correlations between ground truth and synthetic data by time period and sampling frequency tion models based on deep learning because the objective is to advance the prediction in a close time period.In the following sections, this hypothesis is validated; i.e., that the generated data improve the training results of the proposed predictive model.
Table 4
Hyperparameters used for each model.(-) indicates model has no parameter isons as they express the average model prediction error in units of the variable of interest.Their definition is as follows:
Table 5
Results of the MLP technique using ground truth, synthetic, a combination of ground truth + synthetic, ground truth + synthetic with reinforcement learning and shuffled synthetic + ground truth datasets RMSE (root mean square error) MAE (mean absolute error).RMSE and MAE are measured in degrees Celsius ( • C) for each 12 and 24 hours.The best value for each combination of dataset, metric and prediction hour is shown in bold
Table 6
Results of the CNN technique using ground truth, synthetic, a combination of ground truth + synthetic, ground truth + synthetic with reinforcement learning and shuffled synthetic + ground truth datasets RMSE (root mean square error) MAE (mean absolute error).RMSE and MAE are measured in degrees Celsius ( • C) for each 12 and 24 hours.The best value for each combination of dataset, metric and prediction hour is shown in bold
Table 7
Results of the LSTM technique using ground truth, synthetic, a combination of ground truth + synthetic, ground truth + synthetic with reinforcement learning and shuffled synthetic + ground truth datasets RMSE (root mean square error) MAE (mean absolute error).RMSE and MAE are measured in degrees Celsius ( • C) for each 12 and 24 hours.The best value for each combination of dataset, metric and prediction hour is shown in bold
Table 8
Results of the CNN+LSTM technique using ground truth, synthetic, a combination of ground truth + synthetic, ground truth + synthetic with reinforcement learning and shuffled synthetic + ground truth datasets RMSE (root mean square error) MAE (mean absolute error).RMSE and MAE are measured in degrees Celsius ( • C) for each 12 and 24 hours.The best value for each combination of dataset, metric and prediction hour is shown in bold | 8,743 | 2023-07-28T00:00:00.000 | [
"Environmental Science",
"Computer Science",
"Engineering",
"Agricultural and Food Sciences"
] |
Threat and Anxiety in the Climate Debate:An Agent-Based Model to investigate Climate Scepticism and Pro-Environmental Behaviour
How people react to threatening information such as climate change is a complicated matter. While people with a high environmental self-identity tend to react approach-motivated by engaging in pro-environmental behaviour, people of low environmental self-identity may exhibit proximal defence behaviour, by avoiding and distracting themselves from potentially threatening stimuli caused by identified anxious thoughts and circumstances. This psychological theory has recently been tested in experimental studies in which the results suggest that the promotion of climate change information can also backfire. Based on these findings, we propose an agent-based model to address influences on anxiety and correlated pro-environmental actions in relation to societal attitudes of climate change scepticism and environmental self-identity.
Introduction
At a time when the worrying consequences of climate change can no longer be ignored, global warming has become a widely discussed issue.In response to the causal link between current climate change and man-made GHG emissions, environmental organisations, governments and involved stakeholders are trying to motivate citizens to adopt more environmentally friendly lifestyles by presenting information on the consequences of climate change.Although necessary steps to achieve a smaller footprint are well-known, the realisation of a greener lifestyle is proceeding slowly and has been facing stagnation for decades.There is currently no clear answer on how to accelerate the pace of this vital transition to avoid immediate threats to societies.
Despite the profound scientific findings on climate change, public and media discussion is often distorted and shifted to a discussion about whether the scientific facts are valid at all (Dunlap, 2013).While this shift in discussion bias is criticised by climate scientists among several voices, it is often unintentionally supported by failing to take sufficient account of the socio-psychological reasons that can lead to denial of climate change.Calling for social transformation towards a climate-friendly lifestyle through the dissemination of threatening information seems to be based on an intuitively correct assumption: more awareness of the problem of climate change also leads to more climate-friendly actions.The findings of threat and defence research indicate, however, that this approach can create unplanned effects and may also backfire (Uhl, Jonas, and Klackl, 2016).
Here we present an agent-based model of climate communication and the associated climate scepticism and pro-environmental behaviour, using the perspective of a socio-psychological model of threat and defence.Our main goal is to evaluate the model dynamics to identify conditions which can support favourable scenarios and to examine circumstances which enhance or dampen possible backfiring effects.
Threat and Defence Model: Theoretical Background
The social-psychological threat literature deals with reactions to problems such as personal uncertainty, loss of control, conflicting goals or perceptual surprises.Recently, an 'integrative general model of threat and defence processes' was developed by Jonas, McGregor, Klackl, Agroskin, Fritsche, Holbrook, Nash, Proulx, and Quirin (2014) that provides a conceptual framework for understanding such diverse phenomena.Based on societal psychological and neural perspectives on defensive reactions to threat, the model proposes the simple hypothesis that discrepancies arouse anxiety and thereby motivate diverse phenomena that activate approach-related states that can relieve anxiety.
From this perspective, threats result from an experience of the discrepancy between an expectation or a desire and the current circumstances.This discrepancy is followed by anxiety, which leads to a variety of proximal defence reactions such as avoiding the problem.The threat-related processing is mediated by the Behavioural Inhibition System (BIS), which responds with symptoms such as anxiety and avoidance.
In case of potentially threatening information, individuals increase their efforts to suppress or distract and distance themselves from identified anxious thoughts or circumstances.
A second way of combat anxiety is to turn to approach-motivated behaviour.This reaction pattern manifests itself through the Behavioural Approach System (BAS).
When activated, the possible responses include various strategies to seek an effective solution to the problem at hand.Activation is preferred when the discrepancy appears manageable.Since approach-motivated states are able to dampen anxiety and conflict, the anxious BIS stage is successively supplanted or shortened.In case of no available solutions to the threat (e.g., impending death due to ongoing sickness) individuals can nevertheless turn to approach-motivated states by indirectly solving the threat through soothing or mellowing reactive patterns.
In summary, the research of Jonas et al. (2014) suggests that people tend to evade threats over which they feel they have no control and try to relieve anxiety in a symbolic way by turning to more rewarding aspects of life, even if this aspects are unrelated to the actual threat at hand or its solution.However, in the course of time, most people eventually succeed in muting the BIS by engaging in distal approach-oriented reactions, regaining stability and overcoming this negative state.The BAS can be empowered by the pursuit of personal goals by providing a target for the approach-motivated behaviour.Personal goals can be derived from internal orientations (self-identity, values) or social norms.
Climate Change Information
How individuals process negative information about climate change is strongly influenced by their individual beliefs.By visiting several research efforts in the field of environmental psychology that investigated responses to environmental threats, we identify two significant characteristics that are reported as important drivers: (i) climate change scepticism (CCS) and (ii) biospheric values and environmental self-identity (ESI) 1 .Scepticism towards climate change correlates with the belief in a just, orderly and stable world.As a consequence, people with high CCS show only little intention of reducing their environmental footprint when confronted with threatening news (Feinberg and Willer, 2011).On the contrary, these information seem less convincing (Corner, Whitmarsh, and Xenias, 2012).People who are less sceptical are positively influenced in their environmental attitude when confronted with the same information.It has been shown that a high environmental self-identity (i.e. the extent to which you see yourself as a type of person who acts environmentally-friendly) increases pro-environmental behaviour (PEB) of people to whom negative environmental information is presented (Bolderdijk, Gorsira, Keizer, and Steg, 2013).
A recent study examined the responses to a threat exposure by climate information (Uhl et al., 2016;Uhl, Klackl, Hansen, and Jonas, 2018) with reference to 1 Values are general and abstract principles that one strives for in life, while self-identity reflects how one sees oneself.We will limit further discussion to the latter terminology.See (Van der Werff, Steg, and Keizer, 2013) for a conceptual differentiation between environmental preferences, intentions and behaviour.
the threat and defence model (Jonas et al., 2014).Their findings indicate that strong environmental values can not only increase approach-motivated behaviour, but also promote symbolic reactions and can backfire by less willingness to take action, e.g.donate money to an environmental organization.Here, high ESI individuals resolve the threat partly by exhibiting higher PEB intentions, but the same individuals show symbolic defensive behaviour by looking more negatively at multiple groups, including criminals, overweight, or unattractive people.One possible explanation for this is that a higher ESI leads to a greater perceived threat to these participants, who are then unable to fully resolve the threat through purely direct behaviour.
Important work on climate scepticism in a communication network perspective to predict climate change attitudes is presented in (Leombruni, 2015).
We like to contribute to the line of socio-psychological ABM research with a attitude-based model of climate communication.Individual agents have internal orientations that shape their response and the associated probability of environmental behaviour in confrontation with information about climate change.The agent architecture is based on the aforementioned threat and defence model (Jonas et al., 2014) and is backed by findings in (Uhl et al., 2016,1).In addition to response mechanisms to negative information, social contagion processes were implemented via interaction network topologies, for which we use scale-free and spatial-proximity networks to implement agent-to-agent communication.The model dynamics describe the temporal development of internal states of the agents and correlation of environmental actions.
Agent Architecture
We have implemented an internal, attitude-oriented state in connection with communication on climate change.We use the agent architecture (see Table 1) as a composition of an anxiety state (anx) induced by threatening information (e.g. about climate change) and internal orientations of climate change scepticism (ccs) and environmental self-identity (esi).When an agent receives information, an anxiety reaction ∆anx is triggered Fig. 1 (left).The amount of anxiety increase depends on the value of an individual's ccs (colour code) and the information impact II (x-axis), i.e. how severe or negative the information is.Without detailed knowledge of the correlation of stimuli and reactions, we assume linearity.This linear response is added to the anxiety state anx of the individual, which increases most for low ccs and high information impact.If no information is received, the anxiety state follows a small natural rate of decay.With regard to decision making, we implement a pro-environmental behaviour (peb) by using a probability to engage in environmental friendly actions.This probability can increase or decrease according to the values of esi and anx of an individual.The higher these parameters, the more likely an agent takes action.If an agent engages in a peb action, its anxiety is partly released which is linked to a decrease in peb probability.The correlation of the three parameters esi, anx, peb is given by a Hill function as shown in Fig. 1 (right) for different anxiety states.
To facilitate notation, we use lowercase letters when referring to parameter of single agents (esi, ccs, anx, peb) and capital letters when referring to collective means of the population (ESI, CCS, AN X, P EB).The mean of the initial distributions are denoted by ESIinit, CCSinit (see Table 2).The collective orientations or bias of the population ESI, CCS are initially distributed randomly with ESIinit ± 0.2, CCSinit ± 0.2.We would like to point out that personal traits eventually have statistical correlations, so that there is a natural interdependence.In this generic approach we do not concern ourselves with the relationship between the two attributes of ESI and CCS and therefore treat the distributions as independent from each other.
External Information
In addition to documentaries, campaigns, journalistic articles and mouth-to-mouth communication as tools to mobilise public support and educate about climate change, online communication on climate change and climate politics has become increasingly popular and powerful.Although climate scientists and scientific institutions are eager to participate, they do not seem to be major players in online debates (Schäfer, 2012) and their impact on a broader public appears to be limited so far.The problem of effective communication to mobilise citizens to tackle climate change is inherent and reinforced by several factors (Cox, 2010).Since the focus of this study is on social-psychological response mechanisms to threat, a generic version of information streams is sufficient for the implementation of climate-related information.
We implement possible information streams as unified 'external information' given by the information density.Individuals are exposed to negative information, which can vary in severity and frequency.The severity can range from strong to light and is regulated by the information impact II.The information rate IR indicates the probability that an agent receives information in a time step.These two information parameters IR, II control the information density so that we can create scenarios ranging from 'mild but frequent' to 'intensive but sparse' exposure.
Social interactions
The shaping of public opinion through social interactions stems from a complex interplay between mental and social dynamics.This allows attitudes or opinions to emerge, spread and change on a population level.Agent-based modelling is particularly suitable for the study of dynamics involving heterogeneity and social contagion processes (Alvarez-Galvez, 2016;Nardin, Andrighetto, Conte, Székely, Anzola, Elsenbroich, Lotzmann, Neumann, Punzo, and Troitzsch, 2016;Schweitzer and Garcia, 2010;Sopha, Klöckner, and Hertwich, 2013;Tang, Wu, Yu, and Bao, 2015).We use networked ABM, in which interactions are based on an underlying network topology within a population N .We identify climate scepticism as the main contagious attitude with respect to social dynamics, and anxiety as a motivational impulse to participate in possible interactions about climate change.Thus, we model social contagion dynamics of climate change scepticism with a correlation to the internal anxiety level.:The higher the personal anxiety the more likely it is that an agent will communicate and influence one of its link-neighbours on climate scepticism.If an agent is more likely to believe in climate change CCS < 0.5, it can reduce the scepticism of one of its link-neighbours up to a maximal learning rate.On the other hand, if an agent denies climate change to a certain degree CCS > 0.5, that agent can increase the scepticism of a link-neighbour.
The model is limited to CCS contagion, while contagion processes of ESI and P EB are not considered.We believe that climate scepticism is the most contagious attitude for several reasons: First, in relation to the everlasting movement of climate change denial, the contagious effect of CCS has been shown to be highly influential among (some) members of society.Secondly, we believe that the denial of climate change is closely linked to the information received and less to a real and scientific understanding of the situation, which makes this dynamic even faster.Thirdly, environmental self-identity, which involves the process of forming one's own identity, has a stronger need for experience.Therefore, we believe that the time scale of CCS contagion processes is considerably faster and that social contagion processes of ESI are to some extent negligible for our modelling requirements.
The underlying network topology are scale-free networks that exhibit a distribution of degrees (i.e.number of links for each node) that follows a power law.We generated this type of topology using the preferential-attachment algorithm by (Barabási and Albert, 1999).In a second set of investigations to test the robustness of our results, spatial-proximity networks are used.These networks have a high clustering with the number of links in their spatial neighbourhood being given by the average degree d as introduced in (Stonedahl and Wilensky, 2008) to model spreading dynamics of epidemics (SIR model).
Simulations
We use the presented model in order to investigate different aspects of climate communication under ongoing external information flow and social contagion processes over successive rounds t.We do the following simulations using parameters presented in Table 2.
All simulations show temporal dynamics in CCS, AN X, and P EB while ESI and the information density parameters II, IR are fixed for each run.
Simulations for the purpose of calculating final states have a time range of T = 2000 − 10000 plus 500 time steps.The last 500 steps are used for calculating the equilibrium as mean value.The respective simulated total time T depends on the occurrence of an equilibrium.The equilibrium state is reached when the CCS contagion process is complete and the variations of the collective CCS are less than 10 −3 .For almost all simulations, we found equilibration before T = 10000.In the extremely rare case that a simulation did not equilibrate before T = 10000, we examined the results individually and found that very small fluctuations are slightly above our chosen equilibration criterion but no trend in the CCS development was visible for the last 5000 time steps.Simulations for the representation of the temporal development are shortened and serve only the illustration of the dynamics.
Results
Our aim is to compare favourable scenarios of high environmentally friendly behaviour with unfavourable scenarios of low environmentally friendly behaviour.We want to emphasize the conditions that are necessary for pro-environmental behaviour.
For this, we present different analyses of our model, which can be divided into studies on 1. the temporal development of the collective state, 2. the equilibrium results of the collective state, and 3. the heterogeneity of behaviour in the population.
In order to illustrate the influence of the flow of information, we compare scenarios with low, medium and high information density.We use a rigorous parameter analysis of the orientations ESI and CCS, which represent the bias of the population on climate change.This allows us to compare the behaviour of populations of weak to strong environmental self-identity and low to high climate change scepticism.We compare these results on two different interaction topologies (scale-free networks and spatial-proximity networks).To give more insight into trends within a population, we compare the behaviour of climate change 'believers' and climate change 'deniers' as a sub-investigation of the parameter analysis of the internal orientations.We discuss the main effects of the information density on the dynamics and some minor effects regarding the temporal development and similarities under 'mild but frequent' to 'intensive but sparse' exposure.
To simplify the understanding of the results we distinguish the parameter space of
Temporal Evolution
Each simulations starts with no anxiety AN X = 0 and, therefore, no probability to perform P EB.The continuing exposure to information of rate IR and impact II increases the anxiety level which is then followed by a P EB response.This response is delayed due to the necessity of a certain anxiety level to emerge within the population.
The initial mean distribution in the scepticism CCS transforms over time due to interactions of individuals, which is delayed for the same reason than the P EB response.
Fig. 2 shows this processes for a balanced population under exposure of medium information density: First, an increase in anxiety is visible while the P EB probability and CCS distribution are steady.With a delay, the CCS normal distribution is dispersed and strong believers (CCS < 0.1) and strong deniers (CCS > 0.9) become more frequent while the rest of the opinion range is uniformly present.At t = 2000, an equilibrium state is reached with rather high anxiety but lower P EB probability and the population is divided in two unequal groups of strong believers and strong deniers.
The evolution of the collective anxiety level and the correlating P EB actions are shown in Fig. 3 for low, medium and high information density.The results for 20 simulations are given by mean (line) and standard deviations (shaded areas, only visible for P EB at high density).The balanced population with mESI = mCCS = 0.5 is investigated.Comparing these results, the anxiety level in the medium and high density scenario are similar while the correlated approach-motivated behaviour at high density is more than doubled for the medium density case.At low density, an increase in anxiety level is observed but no approach-motivated behaviour occurs.
Similar results can be obtained by using smaller populations N = 1000 (see appendix Fig. 7).No major scaling effects by variations in population size have been observed.
Dependence on Environmental Self-Identity and Climate Change Scepticism
The initial configuration of the population, i.e. whether it is a population with low or high environmental values or climate scepticism (four quadrants), is decisive for the outcome of the behaviour patterns.To test the dependence of the P EB probability on the internal parameters ESI and CCS of the agents, we perform a complete parameter variation, as shown in Fig. 4.Here the initial distributions of both parameters are varied with ESI ∈ [0.01, 1] and CCSinit ∈ [0.01, 1].The results are given for the collective anxiety AN X (left column, Fig. 4) and the collective P EB probability (right column, Fig. 4).The initial mean parameters mESI, mCCS are reflected by the x, y-axis.Results are displayed for the mean of 10 (high) and 5 (low) simulation runs for each set of parameters.
In the case of low information density (top), the collective anxiety is increased in quadrant C, while the other quadrants exhibit rather low anxiety levels.Moreover, for all quadrants the correlated P EB probability is not significant (≈ 0.006) with only a minor increase of 0.012 in quadrant D for populations of very high environmental self-identity ESI > 0.9 (not visible in the colour code).
In the case of high information density (bottom), a correlation of the collective anxiety to the collective CCS value is visible.Populations of high climate scepticism in quadrant A, B (CCS > 0.5) exhibit low anxiety.Populations of higher believe in climate change of quadrant C, D (CCS < 0.5) show higher anxiety levels, with a stronger increase in quadrant C.However, we only observe a positive correlation of the anxiety level and favourable P EB increase for populations with high ESI > 0.5 in quadrant D. Populations of other quadrants show no significant increase in the P EB probability and thus are considered unfavourable.We conclude that for sufficient environmental bias the collective behavioural response with P EB is possible.
The results from Fig. 4 are generated using scale-free networks, which represent a common network topology for societal processes.To test the robustness of our results towards other network types, we explore the same model dynamics on spatial-proximity networks.By using a second topology, we were able to achieve very similar results.
Quadrant A and C do not show P EB responses and quadrant D reacts approach-motivated with increased P EB.In quadrant B an improvement of P EB responses takes place, dividing the quadrant into two areas with and without P EB response, as can be seen in the appendix Fig. 8.
Climate Scepticism Dynamics
Another important aspect of changes in climate scepticism with regard to the amount of information a population is exposed to.To illustrate the social contagion processes on CCS, we examine the relative changes between initial climate scepticism (CCSinit) in a population and the collective equilibrium state CCS, shown in Fig. 5 (left).Values of CCS below the dashed line indicate a decrease in in respect to the initial configuration, while values above CCS indicate an increase in collective CCS.
Interestingly, the higher the information density, the stronger the CCS increase for same initial configurations.Taking the balanced population at CCSinit = 0.5 as an example, we observe a decrease of CCS for all information densities.Inspection of the correlated timelines (not shown) revealed that for the high density case this final decrease in CCS was preceded by a small intermediate increase of CCS ≈ 0.6 which ultimately declined towards CCS ≈ 0.3.
To highlight the correlating P EB actions for these three scenarios, Fig. 5 (right) shows the effect of the information density on the collective P EB.Here, the highest probabilities are given in the case of high information density.We conclude that if the mean initial scepticism is not above 60% the overall positive effect of high information density by enabling agents to increase P EB exceeds the negative effect of higher collective CCS.
Believers and Deniers
The evolution of the collective CCS transforms the initial normal distribution around the initial mean value CCSinit into distinct groups of low and high climate scepticism, as presented in Fig. 2. Since the agent architecture does not include intelligent actions, i.e. agents cannot change their opinions by logical thinking when presented with convincing evidence, this sharp division is a result of the social contagion and a typical phenomenon for generic agent-based modelling without noise.
In order to further investigate this division and its consequences, we categorize the population into two groups.The 'believers' are those whose CCS is below 50%, the 'deniers' are the agents whose CCS is above 50%.We study the parameter dependencies and different behaviour patterns of both groups over the parameter space of internal orientations ESI, CCS (see for comparison Fig. 4).
One of the most important properties is the group size, N B of the believers and N D of the deniers.There is a naturally correlating relationship of the resulting group sizes that is proportional to the initial mean of the normal distribution of scepticism CCSinit.The higher CCSinit the higher the number of N D and vice versa (see Fig. 6 (top left)).Now it is of interest whether this development is affected by the environmental self-identity, especially in the case of high information density, which enables agents to resolve anxiety by responding with P EB actions in quadrant D. This inversion is feasible due to the natural dependence of P EB on the environmental self-identity and verified by the small standard deviations of three out of four curves.
Only the anxiety of the deniers displays both dependencies of ESI and CCSinit.It should also be noted that believers are exclusively contributing to P EB while deniers do not respond with any increase in P EB.Therefore, favourable results of quadrant D (see Fig. 4 (bottom right)) is solely due to the believers, which are motivated to respond with P EB.The deniers can be regarded as free-riders (Heitzig, Lessmann, and Zou, 2011), who benefit from the actions of others without contributing, as they are not willing to contribute.Thus, the group sizes N B , N D are a major influence on the success of the whole population.
In the appendix, this investigation is shown for low information density (Fig. 9).
Additionally, we provide the results on the complete parameter-plane of ESI and CCSinit for both groups of believers and deniers on which this evaluation is based on (low density: Fig. 10, and high density: Fig. 11).
Effects of Information Density
The primary effects of information density encompass a strong influence on the collective anxiety levels and pro-environmental behaviour.However, the positive correlation of favourable P EB progression with increasing information density is only given if certain additional conditions regarding the internal orientations (ESI, CCS) are met (see Fig. 3, 4).Another major aspect is the positive correlation of climate scepticism with increasing information density (see Fig. 5) We would like to add two minor aspects of the impact of information density to the discussion to complete this picture.First, the dynamics of CCS development is slower at low information density than at high.This is caused by a weaker increase in the anxiety levels of individuals and thus a lower probability of participating in communication between agents.Second, with regard to the two density parameters, we have tested whether the information effect or the information rate has a different effect on the results and the balancing time.We tested high rates with low load and the reverse case of high load and low rate.Here we found only a slight difference in the results and runtimes in favor of the information rate.Since the difference is 1%, we do not draw any further conclusions.
Limitations and Outlook
To implement the socio-psychological theory of threat and defence of Jonas et al. (2014) in a networked ABM, we have chose several simplifications to obtain a generic model that can be further extended in future research efforts.Possible extensions or improvements are given in regard to 1) agent architecture and social contagion processes, 2) data on internal orientations (environmental self-identity, climate scepticism), 3) information flow and information processing, and 4) including anxiety release via symbolic defensive behaviour, to name but a few.In order to deal with the questions arising from this work in further detail, we will work on the aspects mentioned below as described below: 1) Further development of the agent architecture is provided, e.g. by the Belief-Obligation-Desire-Intention BOID architecture (Broersen, Dastani, Huang, Hulstijn, and Torre, 2001) where social agents are capable to resolve different conflict types within or among informational and motivational attitudes.
2) Advancing from generic parameter exploration towards more detailed studies, the attribute composition of the population and internal architecture can be improved (distribution, quantity and interrelation) by drawing upon datasets, i.e. from the International Social Survey Programme, Environment Module III (ISSP) provided by the Zentralarchiv für Empirische Sozialforschung, University of Cologne in the GESIS Data Archive (http://zacat.gesis.org/).
2) In addition to social contagion influences on climate scepticism the external information and correlated information processing can impact opinion dynamics within a population.Implementation of a direct influence of scepticism through convincing or persuasive information represents a meaningful extension of the model.
4) The model represents a closed system, where agent's only option to reduce anxiety is by approach-motivated behaviour or small natural decrease when not confronted with negative information.We are looking forward to conceptualize a model version that takes into account the anxiety reduction of the resulting symbolic defensive behaviour that manifests itself, for example, as ethnocentrism or other discriminatory behaviour.
Ultimately, efforts to model communication on climate change should not only help us to understand how related social norms are formed and passed on, but also enrich our knowledge about how to manage the transformation towards a greener lifestyle.With this in mind, we intent to further improve the model capacities in the near future and to provide a supporting tool for critical reflection on the key challenges in connection with climate change mitigation.We would like to emphasize that our understanding of ABM does not include predictive power, but the nature of the model as presented in this work serves the purpose of enrichment of the discourse.
Conclusion
In this study, we investigated the climate debate from the socio-psychological perspective of thread and defence research by developing and analysing an agent-based model.We focus on long-term population-based effects that can be caused by information about climate change.In a nutshell, the dynamics follow a few simple principles: The exposure of threatening information may cause an anxiety response of an individual which then can be released by approach-motivated pro-environmental behaviour.The actual increase in environmentally friendly behaviour is subject to several factors, such as internal orientations on climate change and environmental self-identity, but also on the amount of received information.
In general, the confrontation with the consequences of climate change does not increase the pro-environmental intent unless several conditions are met: 1.The amount of information has to be large enough to encourage approach-motivated action, otherwise the anxiety increases but does not sufficiently promote such behaviour.
2. The majority of the population has a well-developed environmental self-identity, otherwise they lack in motivation to respond with environmental friendly behaviour.
3. The majority of the population should strongly believe in climate change, otherwise free-riding of climate change deniers reduce the overall success of the population.
An important side effect that we observed was in correlation to the amount of information density.We observed a relative increase in public scepticism the more information was provided.This reflects an alternative way of dealing with anxiety by avoiding restless thoughts and moving away from them.Encouragingly, environmentally friendly actions were positively correlated to information increase and this effect more than compensated for the negative effects of increasing scepticism.
We would like to highlight some of our most important insights of the model dynamics: • An increased anxiety about climate change is not necessarily associated with greener lifestyle.
• High information density about climate change is not necessarily correlated with favourable scenarios, while low information density will always lead to unfavourable scenarios.
• The number of climate change believers is important for the success of a population, while climate change deniers are free-riding.
• Believers in climate change of low environmental orientations are highly anxious but refrain from approach-motivated behaviour, which is particularly important in regard to symbolic defensive behaviour (backfiring effect).
• Believers in climate change of high environmental orientation engage in sufficient environment friendly lifestyles only if high amounts of information are continuously provided.
• Initial high values on climate scepticism hinder the population to develop anxiety while the minority of believers are highly anxious but muted by the denying majority.
• Balanced societies without a bias in scepticism evolved towards less scepticism, while this trend was dampened the more information was distributed.
• Initially highly sceptical societies showed trends towards even more scepticism, while this trend was intensified as more information was distributed.
Regarding the non-zero anxiety states for all levels of environmental friendly orientations, our research is in line with the results in Uhl et al. (2016), which showed that both groups of low and high environmental self-identity can show symbolic responding in order to release anxiety.
Drawing a general conclusion on the observed collective behaviour of the modelled population, we find that the conditions under which a transition to greener lifestyles takes place are very narrow.In order to motivate the population, an explicitly high ecological identity alone is not sufficient, but it would also require an ongoing high information load, which, realistically, cannot be achieved in real-world systems.We believe that this indicates that a self-regulated transition, meaning that individuals choosing voluntarily to change behaviour permanently, is rather unlikely.This leads us to conclude that systemic regulation, including environmental laws, sanctions for exceeding footprints, and financial incentives is needed to enable a transition towards sustainable societies.This portrayal is model-based only and might be rather pessimistic, since we did not include several powerful motivators, such as hope in the context of climate change (Chadwick, 2015) or overcoming scepticism, for example through education (Stevenson, Peterson, Bondell, Moore, and Carrier, 2014).
ESI, CCS in four quadrants A, B, C and D. Each quadrant represents populations with a specific bias.Quadrant A: more sceptical and less environment-oriented, quadrant B: more sceptical and more environment-oriented, quadrant C: less sceptical and less environment-oriented, and quadrant D: less sceptical and more environment-oriented.The balanced population of ESI = 0.5 and CCS = 0.5 is centred in-between the quadrants.
Fig. 6 (bottom) shows a similar investigation for believers and deniers and their levels of anxiety and pro-environmental behaviour in relation to both internal orientations.Here the perspective on CC and ESI is inverted: Values with identical CCSinit are averaged and the correlation to ESI is depicted by the x-axis.The mean value (line) and the standard deviation (shaded area) of AN X and P EB are given.
Figure 1 .Figure 2 .Figure 3 .Figure 4 .Figure 6 .Figure 7 .Figure 8 .Figure 10 .
Figure 1 .Agent architecture: (left) Anxiety response ∆anx in relation to the information impact shown for six different ccs values; (right) agents peb probability in dependence of the environmental self-identity esi shown for different six different anxiety states anx.
Table 1
Agent internal state and attitudes: parameters, ranges and explanations.Model parameters, ranges and explanations of the full set of simulation parameters. | 7,751.8 | 2019-07-10T00:00:00.000 | [
"Environmental Science",
"Political Science",
"Philosophy"
] |
Interplay Among Hydrogen Sulfide, Nitric Oxide, Reactive Oxygen Species, and Mitochondrial DNA Oxidative Damage
Hydrogen sulfide (H2S), nitric oxide (NO), and reactive oxygen species (ROS) play essential signaling roles in cells by oxidative post-translational modification within suitable ranges of concentration. All of them contribute to the balance of redox and are involved in the DNA damage and repair pathways. However, the damage and repair pathways of mitochondrial DNA (mtDNA) are complicated, and the interactions among NO, H2S, ROS, and mtDNA damage are also intricate. This article summarized the current knowledge about the metabolism of H2S, NO, and ROS and their roles in maintaining redox balance and regulating the repair pathway of mtDNA damage in plants. The three reactive species may likely influence each other in their generation, elimination, and signaling actions, indicating a crosstalk relationship between them. In addition, NO and H2S are reported to be involved in epigenetic variations by participating in various cell metabolisms, including (nuclear and mitochondrial) DNA damage and repair. Nevertheless, the research on the details of NO and H2S in regulating DNA damage repair of plants is in its infancy, especially in mtDNA.
INTRODUCTION
Hydrogen sulfide (H 2 S), nitric oxide (NO), and reactive oxygen species (ROS) including superoxide anion (O 2 · − ), hydroxyl radical (HO·), and hydrogen peroxide (H 2 O 2 ) are important intercellular signaling agents in living organisms due to their high activity, small size, and high membrane permeability (Porrini et al., 2020). For instance, H 2 O 2 acts as an essential second messenger in the oxidative reactions with cysteine residues, producing the post-translationally modified proteins potent to redox signaling (Forman et al., 2010). H 2 S and NO perform their versatile roles in plants primarily due to the protein S-nitrosation and persulfidation, respectively, which are oxidative post-translational modifications of cysteine residues (Filipovic and Jovanovic, 2017;Porrini et al., 2020). NO can also react with O 2 · − and H 2 S to produce signaling molecules peroxynitrite and S-nitrothiols, respectively (Hancock, 2017). In plants, all these endogenously generated reactive species appear to play multiple roles in many crucial physiological and biochemical processes (Figure 1), including modulating seed germination, maintaining plant growth and development, regulating plant senescence and fruit ripening, and improving the tolerance to biotic or abiotic stresses (Corpas, 2019b;Huang et al., 2019). The positive effects of H 2 S, NO, and ROS depend on their proper concentrations; if not, the toxic effects show up. So, learning the exact boundary between the physiological and toxicological concentration of endogenous ROS, NO, and H 2 S contents is essential. The in vivo concentration of these three active substances showed great differences when measured via different methods. The entire H 2 S content ranges from 0.010 to 0.199 µmol g −1 FW by the methylene blue method. Using the electrode method, the H 2 S content showed a wider range, from 0.177 to 0.708 µmol g −1 FW (Jin et al., 2018). Low-level H 2 S and NO have acted as signaling molecules to delay the senescence of postharvest fruits and ameliorate cold stress injuries (Geng et al., 2019;Wang et al., 2021). When their concentrations reach higher levels than normal, stresses are sensed. The excessive reactive species constantly attack biomolecules, leading to severe or even irreversible oxidative modification, such as lipid peroxidation, protein oxidation, and oxidative DNA damage, which can induce membrane damage, functional changes, strand breaks, and even lead to cell death (Trachootham et al., 2008). The uncontrolled increasing level of ROS generated in plants under stress conditions often induces abnormal growth or even death of plants . Excessive H 2 S negatively affects the mitochondrial respiratory chain with the inhibition of cytochrome c oxidase by redox-reacting with the metal center (Filipovic et al., 2018).
Therefore, given the dual role depending on concentrations, maintaining the balance between production and elimination of H 2 S, NO, and ROS in plants is pivotal to continue their positive roles. The abiotic stresses elevate ROS production and may concomitantly accumulate endogenous NO and H 2 S, which can reduce oxidative stress by promoting antioxidative defenses to scavenge ROS (Bhuyan et al., 2020). Moreover, excessive ROS may convert NO to ONOO − , which triggers nitrooxidative stress (Corpas and Barroso, 2013), leading to protein tyrosine nitration. H 2 S can increase the endogenous NO level (Amooaghaie et al., 2017;Corpas et al., 2019), and NO can likewise induce the accumulation of endogenous H 2 S in plants (Peng et al., 2016). Taken together, as bioactive species, ROS, NO, and H 2 S independently or collaboratively act with each other to participate in the regulation of diverse cellular processes. In this study, we summarized the production and elimination of ROS, NO, and H 2 S in plants and their roles in cellular redox balance and DNA damage and repair, attempting to describe the interplay between these reactive species from their metabolism to their regulatory performance in responding to stresses. On top of that, we described the prospects of NO, H 2 S, and ROS in the field of mitochondria and mitochondrial DNA (mtDNA) in plants.
METABOLISM OF ROS, NO, AND H 2 S IN PLANTS
Generally, ROS can be produced in the cellular compartments of plants from aerobic metabolism, including respiration and photosynthesis, and regulate the life processes of plants, exhibiting its positive or toxic roles under normal or abnormal conditions (Jiao et al., 2014;Yu et al., 2017;Waszczak et al., 2018). In higher plants, ROS can be produced in chloroplasts, mitochondria, cell membranes, and peroxisomes, among which the chloroplasts and mitochondria are the main production sites and produce ROS primarily through the electron transport chain (ETC) (Fluhr, 2009). In mitochondria of plants, the ETC includes the cytochrome pathway and alternative pathway, but H 2 O 2 or O 2 − is not the end product of the reduction of oxygen in the alternate pathway (Moore and Siedow, 1991). The respiratory complexes I and III are the major sites of O 2 ·− production in mitochondria. O 2 ·− can be dismutated spontaneously or catalyzed by mitochondrial manganese-superoxide dismutase (Mn-SOD) into H 2 O 2 , which is then detoxified by peroxiredoxin (Prx), glutathione peroxidase (GPX), and ascorbate peroxidase (APX) of the ascorbate-glutathione cycle in plant mitochondria (Jimenez et al., 1997;Chew et al., 2003). H 2 O 2 can be converted into a hydroxy radical (HO) when reacted with redox-active transition metals. Additional endogenous sources of ROS comprise the membrane-associated NAD(P)H oxidase (NOX) and xanthine oxidase (XO) (Trachootham et al., 2008). Studies showed that NOXs could regulate plant cell expansion by producing ROS, and the knockout of RHD2, an NADPH oxidase, resulted in decreased ROS content and short root hairs and stunted roots in Arabidopsis (Foreman et al., 2003). Besides, NADPH-mediated ROS generation was found to be related to the stomatal closing (Kwak et al., 2003), further indicating the important functions of ROS in plant growth and development. Food extracts such as hesperetin could inhibit the activity of XO, reducing oxidative stress caused by ROS (Dew et al., 2005). ROS can be scavenged via enzymatic antioxidant systems mainly including SOD, catalase (CAT), peroxidase (POD), APX, GPX, monodehydroascorbate reductase (MDHAR), dehydroascorbate reductase (DHAR), and glutathione reductase (GR), and nonenzymatic antioxidants, including NAD(P)H, glutathione (GSH), ascorbic acid (AsA), and flavonoids, which is well summarized by Huang et al. (2019).
Nitric oxide, as a small and redox-active molecule, plays versatile roles in the physiological and biochemical processes in plants (Brouquisse, 2019;Gupta et al., 2020). Dynamic monitoring of NO in the imbibed seeds showed that the amount of NO released reached a peak at 3 h; moreover, increased NO release (5 nmol min −1 g −1 DW) observed in Arabidopsis seeds treated with sodium nitroprusside (SNP, an NO donor) enhanced seed germination (Liu et al., 2009). For a long time, researchers dedicated themselves to clone nitric oxide synthase (NOS) from plants to figure out the biosynthetic pathway of NO but failed. Reports showed that NO can be produced from nitrite through the nonenzymatic reduction and the catalysis by nitrite reductase, nitrate reductase, or molybdoenzymes and from an oxidative route with NOS-like activity in plants (Astier et al., 2018;Kolbert et al., 2019). The mitochondrial electron transport chain (mETC) also involves the production of NO from nitrite (Alber et al., 2017;Gupta et al., 2018). Complex I (NADH: ubiquinone oxidoreductase), cooperating with rotenone-insensitive NAD(P)H dehydrogenases, regulates the production of NO under hypoxia; Complex II (succinate: ubiquinone oxidoreductase) contains Fe-S centers which can be inhibited by NO, regulating the generation of ROS; Complex III (ubiquinol: cytochrome c oxidoreductase) transfers an electron to nitrite to generate NO; Complex IV (cytochrome c oxidase) also contributes to the interconversion between nitrite and NO and the generation of ATP via the phytoglobin-NO cycle (Alber et al., 2017;Gupta et al., 2018). Complex V (F 1 F O -ATPase) synthesizes most of the ATP and provides energy to living cells. Tyrosine nitration caused by NO can inhibit the F 1 F O -ATPase activity (Nesci et al., 2017). Many studies have shown that excess NO can inhibit oxygen metabolism by the cytochrome pathway; however, the inhibition of oxygen metabolism in the alternative pathway is very weak (Millar and Day, 1996). The cDNA microarray and Northern analysis evidenced that appropriate NO levels could induce the transcription of alternative oxidase (AOX) in Arabidopsis cells (Huang et al., 2002). Meanwhile, the activation of NO on AOX transcription was also observed in tobacco (Ederli et al., 2006), and the relationship between NO and AOX was further explored by Cvetkovska and Vanlerberghe (2012) by using the transgenic material of AOX in tobacco, and a preliminary conclusion was drawn: AOX respiration acts to reduce the generation of ROS and reactive nitrogen species (RNS) in plant mitochondria by dampening the leak of a single electron from the ETC to O 2 or nitrite. Excess NO may also trigger the S-nitrosation of complex I and other proteins.
Hydrogen sulfide has been identified as a new endogenous player in plants, and it is a mitochondrial substrate or a poison to mitochondria at low or high concentrations, respectively . The role of H 2 S in cells has long been a matter of controversy: Hancock and Whiteman (2014) categorize H 2 S as a referee, one for its role in responding to stresses by interacting with NO and ROS metabolism passively and another for the lack of a dedicated pathway in which H 2 S responds to such stresses. H 2 S in plants is more often considered to be a signaling molecule for its important and irreplaceable function in various physiological processes of plant cells. In plants, the diversity of sources ensured the needed supplement of H 2 S. H 2 S comes from the environment and the endogenous generation via enzymatic and nonenzymatic pathways in chloroplasts, mitochondria, and cytosols. Plants actively take up H 2 S in the atmosphere via the foliage, which can be converted into GSH in cells, leading to the increased accumulation of thiol, consequently (Ausma and De Kok, 2019). H 2 S can be generated in plants through various biosynthetic pathways, such as sulfite reductase (SiR): converting sulfite to H 2 S, cysteine desulfhydrase (CD): converting cysteine to pyruvate and H 2 S, and cysteine synthase (CS): converting L-cysteine to O-acetyl-L-serine and H 2 S. Mitochondrial β-cyanoalanine synthase (β-CAS) catalyzes the conversion of cyanide and L-cysteine to β-cyanoalanine producing H 2 S (Gotor et al., 2019). Nitrogenase Fe-S clusters also contribute to the generation of H 2 S from L-cysteine in mitochondria . H 2 S can be detoxicated by sulfide: quinone oxidoreductase (SQR), superoxide dismutase (SOD), or O-acetylserine(thiol)lyase C (OAS-TL C). SQR converts H 2 S to persulfides, which can be transferred into the conversion between GSH and glutathione persulfide (GSSH), and the sulfur is then sequentially oxidized by the mitochondrial sulfur dioxygenase ETHE1 to sulfite and by sulfite oxidase (SO) to sulfate or thiosulfate regenerated by thiosulfate sulfurtransferase (TST), transferring sulfur from GSSH to sulfite (Olson, 2018). Mitochondrial Mn-SOD also catalyzes the oxidation of H 2 S into polysulfides (D'Imprima et al., 2016;Olson, 2018). OAS-TL C could transfer sulfide to O-acetylserine to form cysteine in the mitochondria of Arabidopsis (Birke et al., 2015).
H 2 S, NO, AND REDOX BALANCE IN PLANTS
Reactive oxygen species, reactive sulfur species (RSS), and RNS burst in plants under both biotic and abiotic stress could cause the plants to responds to abnormal conditions (Jia et al., 2016;Yu et al., 2017;Corpas, 2019a;Hancock, 2019) and participate in the ripening of fruit (Muñoz-Vargas et al., 2020). During the ripening and stress response processes, ROS accumulation is ubiquitous in several sub-organelles of the cell, such as mitochondria, chloroplasts, and peroxisomes. Accompanied with this, intracellular RNS, such as NO, also increased, which alleviated oxidative damage caused by excessive accumulation of ROS via the intracellular antioxidant pathway . The increasing intracellular NO could also delay the maturation and senescence of fruits by regulating ethylene biosynthesis and the ethylene signal transduction pathway and by affecting the genes of cell wall degrading enzymes (Lin et al., 2020). Mutation of SlLCD1, the H 2 S-producing enzyme, lowered the H 2 S content and accelerated fruit ripening in tomato, while exogenous H 2 S treatment in unripe fruits of tomato suppressed the expression of the ripening-related gene, suggesting that H 2 S is involved in the regulation of fruit ripening (Hu et al., 2020).
Pretreatment with exogenous H 2 S could upregulate the activities of antioxidant enzymes to remove excessive ROS and reduce oxidative damage in Chinese cabbage roots, therefore alleviating the growth inhibition caused by cadmium . Exogenous H 2 S could decrease mitochondrial permeability transition and ROS contents but increase mitochondrial membrane fluidity, mitochondrial membrane potential, and antioxidative enzyme activities in roots of Malus hupehensis under NaCl stress (Wei et al., 2019). The salinityinduced augmentation of H 2 S and NO levels is associated with the increase of L-cysteine and L-arginine and the induction of the enzymes involved in the biosynthesis of H 2 S or NO (da Silva et al., 2017). Sodium hydrosulfide (NaHS) alleviates oxidative damage by increasing the activities of SOD, CAT, POD, and APX, promoting the transcript level of CsNMAPK and the accumulation of endogenous NO through the MAPK/NO signal pathway in cucumber against excess nitrate stress (Qi et al., 2019). The enhanced accumulation of GSH induced by H 2 S alters the redox of the cell, which may consequently increase the tolerance of plants to environmental stress (Noctor et al., 2012).
Both NaHS and SNP can increase the endogenous NO level and enhance the antioxidant enzyme activities in Bermuda grass under lead stress (Amooaghaie et al., 2017). Exogenous NaHS increases the content of endogenous H 2 S and the activity of L-cysteine desulfhydrases (L-CD) in tomatoes under nitrate stress and induces the synthesis of NO through nitrate reductase and not NOS (Liang et al., 2018). Exogenous SNP could also induce the accumulation of endogenous H 2 S by increasing the activities of L-CD, OAS-TL, and β-CAS, and exogenous NaHS enhances the NO-induced hypoxia tolerance in maize (Peng et al., 2016). However, exogenous NO decreases the activities of L-/D-CD, OAS-TL, and SiR and increases the activity of β-CAS, leading to a decrease in the contents of endogenous H 2 S, cysteine, and sulfite in peaches during cold storage (Geng et al., 2019). Treatment of exogenous H 2 S at high concentration triggered the activation of MPK6, consequently inducing the production of NO, and in turn, the exogenous H 2 S-mediated changes in auxin distribution were regulated by NO produced here, resulting in the inhibition of the primary root growth in Arabidopsis . cPTIO (as an NO scavenger) and sodium tungstate (as an inhibitor of nitrate reductase) increase the H 2 S content by sustaining the activities of L-/D-CDs, OAS-TL, and SiR; L-NAME (as an inhibitor of NOS-like activity) improves the H 2 S content mainly by maintaining the D-CD activity, suggesting that there would be different interactions between the NO biosynthesis and the H 2 S metabolism (Geng et al., 2019). Also, hypotaurine (as the specific scavenger of H 2 S) reduces the endogenous H 2 S levels and reverses the responses induced by NaHS but cannot completely reverse the responses induced by SNP; while cPTIO quenches the effects of both NaHS and SNP on Pb tolerance of Sesamum (Amooaghaie et al., 2017). These results also suggest that there would be a two-sided link between the signal molecules of H 2 S and NO (Figure 2) and that H 2 S increases the NO production and, subsequently, NO tightly regulates feedback of the H 2 S biosynthesis (Amooaghaie et al., 2017). The complicated FIGURE 2 | A schematic graph of the crosstalks among H 2 S, NO, and ROS in their metabolisms and regulation on DNA oxidative damage and repair. The booming ROS could trigger the accumulation of H 2 S and NO by promoting the activities of enzymes involved in their bioproduction under abnormal conditions. Both H 2 S and NO could reduce oxidative stress by inhibiting the production of ROS, promoting antioxidative defenses to scavenge excessive ROS, and mutually elevating their production. Both H 2 S and NO mediate the repair of DNA oxidative damage caused by ROS. AOX, alternative oxidase; APX, ascorbate peroxidase; AsA, ascorbic acid; CAT, catalase; CAS, cyanoalanine synthase; CD, cysteine desulfhydrase; CS, cysteine synthase; COX, cytochrome c oxidase; DHAR, dehydroascorbate reductase; GSH, glutathione; GR, glutathione reductase; GPX, glutathione peroxidase; mETC, mitochondrial electron transport chain; MDHAR, monodehydroascorbate reductase; NR, nitrate reductase; NOX, NAD(P)H oxidase; NOS-like, nitric oxide synthase like; OAS-TL C, O-acetylserine(thiol)lyase C; POD, peroxidase; Prx, peroxiredoxin; SQR, sulfide: quinone oxidoreductase; SO, sulfite oxidase; SiR, sulfite reductase; SOD, superoxide dismutase; TST, thiosulfate sulfurtransferase; XO, xanthine oxidase. relationships between the H 2 S and the NO signaling cascade might depend on their respective concentrations, the different physiological and biochemical processes, different tissues and organs, and different species of plants under normal conditions or varied conditions of abiotic stress.
The current knowledge of the regulation by H 2 S, NO, and ROS on plant defense against abiotic stress has been well reviewed (Bhuyan et al., 2020). Interactions among H 2 S, NO, and ROS in mitochondria, cytoplasm, and chloroplast influenced the responses of the plant to abiotic stresses and the processes of ripening and senescence (Hancock and Whiteman, 2016;Zhang P. et al., 2017;Muñoz-Vargas et al., 2018;González-Gordo et al., 2020). To survive, plants try to maintain a proper cellular redox balance. However, the interactions among H 2 S, NO, and redox balance are still not precise and need to be further improvised.
MITOCHONDRIAL DNA OXIDATIVE DAMAGE AND REPAIR IN PLANTS
The mitochondrion has its own DNA (mtDNA), which is highly variable in size and structure depending on the species. The mtDNA in plants is larger than human mtDNA, typically around 200-400 kb, which can be much bigger, reaching up to 11.3 Mb, and can encode about 20 additional genes in comparison to animals (Chevigny et al., 2020). The mtDNA is located near the electron transfer chain and encodes many critical proteins for the assembly and activity of the mitochondrial respiratory complexes (Alencar et al., 2019). The mtDNA is a circular molecule and is packed by proteins including prohibitins, the ATPase family AAA domain-containing protein 3 (ATAD3), mitochondrial transcription factor A (TFAM), DNA polymerase gamma, catalytic subunit (POLG), etc., forming a nucleoid that uniformly distributes within the mitochondrial matrix, which is essential for mitochondrial functions.
Proper levels of biological oxidations originating in mitochondria fulfill the beneficial roles in redox homeostasis, while excessive oxidants overwhelming the antioxidant defenses cause redox imbalance, disrupt mitochondrial function, and lead to dysfunction, aging, and cell death (Piantadosi, 2020). ROS (especially HO·) have a single electron and are prone to nucleophilic attack on DNA molecules, resulting in changes such as modification of DNA sequences, causing pairing and coding errors during DNA replication, and gene mutations (Poetsch, 2020). The DNA oxidative damage includes base modifications, abasic sites, and strand breaks (Gonzalez-Hunt et al., 2018). The stability of the mtDNA is essential for proper mitochondrial function. However, the mtDNA can more easily be injured by ROS on account of its proximity to the site of ROS generation and the impotent mtDNA repair (Trachootham et al., 2008). For example, exogenous H 2 O 2 treatment resulted in the oxidation of Polγ and reduction in exonuclease activity, which in turn converted the high fidelity Polγ into an editingdeficient polymerase, leading to an increase in mtDNA mutations (Anderson et al., 2019).
The preservation of DNA integrity is necessary for ensuring unperturbed transcription (Lans et al., 2019). Unlike other cell macromolecules, damaged DNA cannot be replaced and can only maintain its integrity through direct damage reversal, mismatch repair (MMR), nucleotide excision repair (NER), base excision repair (BER), and recombination pathways [targeting DNA damages of double-strand breaks (DSBs) and singlestrand gaps (SSGs)] (Chevigny et al., 2020). Direct damage reversal is the simplest way with the activities of photolyase, alkyltransferase, and dioxygenase to restore the damaged base of cellar DNA without excision of the base or the phosphodiester backbone (Yi and He, 2013). For example, photolyase could use visible light to transfer electrons from FADH − to cyclobutane pyrimidine dimers (CPDs), a major UV-induced lesion in DNA in plants, resulting in CPD splitting (Zhang M. et al., 2017); alkyltransferase could repair O(6)-methylguanine in DNA by transferring the methyl group (Pegg, 2000); and dioxygenase could also demethylate DNA methylation by oxidizing 5methylcytosine (5-meC) (Wu et al., 2018).
The non-canonical base pairing and insertion-deletion loops can be repaired via the MMR pathway, which is usually associated with the replication machinery in the nuclear matrix but is unclear in the mitochondria (Chevigny et al., 2020). The NER pathway repairs the lesions of DNA caused by UV radiation by removing DNA-binding lesions and adducts, creating a gap, which will be filled by synthesizing damage-free DNA by polymerases to finally be ligated by sealing the nick (Kobaisi et al., 2019).
However, NER and MMR pathways do not exist in plant mitochondria (Van Houten et al., 2018;Wynn et al., 2020), and BER is the primary repair pathway for mtDNA oxidative damage (Ferrando et al., 2019). The deaminations, oxidations, alkylations, and single-strand breaks of DNA can be repaired via the BER pathway (Alencar et al., 2019), the glycolytic excision of the damaged base activated the BER, and the diversity of DNA glycosylases that specifically recognize different types of lesions determined the efficiency of BER (Krokan et al., 1997). Uracil-DNA glycosylase (UNG) from Arabidopsis is imported into the mitochondria (Boesch et al., 2009), combined with UNG found in maize and potato (Bensen and Warner, 1987;Ferrando et al., 2019), indicating that UNG is present in the mitochondria of plants and contributes to the repair of the mtDNA. In addition, double-strand break repair (DSBR) is suggested to be a general system of repairing many DNA lesions in plant mitochondria (Wynn et al., 2020). Among the four distinct DSBR pathways, including non-homologous DNA end joining (NHEJ), alternate end joining (a-EJ), homologous recombination (HR), and single-strand annealing (SSA), HR is considered as the primary DNA repair pathway in the mitochondria of plants (Chevigny et al., 2020).
H 2 S AND NO AFFECT mtDNA OXIDATIVE DAMAGE
Oxidative damage of nuclear DNA and mtDNA can be induced by the excessive accumulation of ROS in plant cells, which could also cause epigenetic variations in plants, such as DNA methylation/demethylation (Katsuya-Gaviria et al., 2020;Nagaraja et al., 2021) and histone modifications (Zheng et al., 2021) influencing plant development and growth. DNA damage caused by ROS could trigger the nuclear redox network and affect DNA metabolism through redox-dependent regulatory mechanisms comprising redox buffering and posttranslational modifications, such as the thiol-disulfide switch, glutathionylation, and S-nitrosation (Cimini et al., 2019). Recently, it has been found that ROS can function as catalysts of DNA methylation (Wu and Ni, 2015;Teng et al., 2018). As mentioned before, both NO and H 2 S within moderate concentration could maintain redox balance by involving the ROS metabolism and the antioxidant system, scavenging excess ROS and, thus, mitigating the oxidative DNA damage; moreover, NO and H 2 S were verified to regulate the expression and posttranscriptional modification of proteins related to DNA oxidative damage and repair (Figure 2).
Nitric oxide and H 2 S can be involved in the DNA oxidative damage via epigenetic modification. The most extensively studied and characterized epigenetic modification of DNA is the methylation of cytosine (C) with an addition of a methyl group to carbon 5 (C5) of the pyrimidine ring (5-meC) (Law and Jacobsen, 2009;Michalak et al., 2013). The methylation state of plant genomic DNA will change into the hypermethylation/hypomethylation form to affect the structure of chromatin and DNA conformation, DNA stability, and the way DNA interacts with proteins, as well as the expression of related genes under abiotic stress (Liu et al., 2017;Zhang et al., 2018). Besides, DNA demethylation, mediated by teneleven translocation dioxygenase (TET) 3, is reported to be crucial for efficient repair of DNA damage (Jiang et al., 2017). Studies on the roles of NO and H 2 S in DNA/mtDNA oxidative damage are more advanced in mammals. For example, mtDNA haplogroup J, which was proved to be associated with several multifactorial diseases and aging, modulates NO production (Fernández-Moreno et al., 2011), and people carrying the mtDNA haplogroup J show lower mitochondrial oxidative damage (Martínez-Redondo et al., 2010); reduction of NO and DNA/RNA oxidation products were observed in patients with systemic lupus erythematosus, and NOx levels and DNA/RNA oxidation products were inversely and independently associated (Iriyoda et al., 2017), all of these indicating that NO is associated with DNA oxidative damage. NO takes part in the regulation of DNA methylation, although likely to be genotoxic at high concentrations. Excessive NO can cause the deamination of cytosine to uracil in single-stranded DNA cytosine residues, resulting in DNA/mtDNA damage, histone deamination (Merchant et al., 1996). As an NO donor, SNP at high concentration inhibits the growth of rice seedlings, which is associated with hypomethylation at the CHG sites (H=A, C, or T) of genomic DNA and the transcriptional activation of genes and transposable elements, and the DNA methylation caused by SNP is inherited by the next generation (Ou et al., 2015). Proper concentration of exogenous NO mitigates the increase of genomic template instability, DNA methylation, and retrotransposon polymorphism caused by copper stress by increasing the efficiency of the antioxidative system in lettuce (Yagci et al., 2019). Studies in smooth muscle cells and aorta tissues of mice found that a sufficient level of H 2 S was able to inhibit TFAM promoter methylation and maintain the mtDNA copy number (Ou et al., 2015). Methyl in trans-methylation reactions are tightly coupled with the activated methyl cycle, a crucial contributor to DNA and RNA methylation in stress-exposed plants (Rahikainen et al., 2018). As a donor of the methyl in trans-methylation reactions, S-adenosyll-methionine (SAM) was associated with the production of H 2 S (Eto and Kimura, 2002). NO can regulate enzymes such as S-adenosylhomocysteine hydrolase/homologous gene silencing 1, methionine synthase, and S-adenosyl methionine synthase/methionine adenosyltransferases in SAM synthesis via S-nitrosation and tyrosine nitration (Lindermayr et al., 2005;Kumar et al., 2020), indicating the crosstalk between H 2 S and NO in regulating DNA methylation/demethylation.
On the other hand, NO and H 2 S could participate in the DNA/mtDNA damage (repair) via post-transcriptional modification of proteins. NO-mediated increase in DNAdependent protein kinase catalytic subunit (DNA-PKcs), a key double-strand DNA break repair enzyme involved in nonhomologous end-joining, demonstrated the presence of a new and highly effective NO-mediated mechanism for DNA repair through S-nitrosation and transcriptional regulation (Xu et al., 2000). NO also modifies histone methylation by regulating protein arginine methyltransferase activity by S-nitrosation, upregulating the expression of the gene encoding lysine methyltransferase, which is the predominant mechanism for transduction of NO bioactivity (Hussain et al., 2016;Blanc and Richard, 2017). Similarly, the NO donor treatment resulted in tyrosine nitration and inhibition of its activity possibly through S-nitrosation, which involved DNA repair (Jones et al., 2009). The redox modifications, such as the S-nitrosation caused by NO, may inhibit histone deacetylases (HDAC 2C and 2B) and modulate histone acetylation in Arabidopsis (Chaki et al., 2015;Mengel et al., 2017). Histones, acetyltransferases, and methyltransferases are the targets for persulfidation (Aroca et al., 2018), suggesting that H 2 S can also participate in DNA repair like NO. Exogenous SNP and S-nitrosoglutathione (GSNO), as NO donors, cause the S-nitrosation of Cys49 and Cys53, promoting a conformational change in the secondary structure in proteins of the AtMYB30 transcription factor and inhibiting the DNA binding ability of R2R3-MYB2 from Arabidopsis (Serpa et al., 2007;Tavares et al., 2014). Histone deamination could be repaired through the BER pathway, which is responsible for the repair of damaged single bases resulting from deamination, alkylation, and oxidized bases (Cimini et al., 2019). H 2 S can modify the thiol group of cysteine (-SH) in proteins into a persulfide group (-SSH) through the process of S-sulfhydration, which is considered as the protective mechanism for proteins against oxidative damage (Aroca et al., 2018). Cysteine residues of proteins can be modified through both S-nitrosation by NO and S-sulfhydration by H 2 S, suggesting that cysteine may be a hub between the physiological effects of H 2 S and NO, and the S-nitrosation and S-sulfhydration of cysteine may be interconvertible.
Moreover, studies in rats showed that pretreatment of NaHS attenuated Hcy-induced mitochondrial toxicity caused by excessive ROS and mito-ROS and restored ATP production and mtDNA copy numbers as well as oxygen consumption in the osteoblast (Zhai et al., 2019). NaHS significantly reduced oxidative stress and attenuated the mitochondrial damage induced by methylmercury (MeHg), and they increased DNA and RNA content in the rat cerebral cortex (Han et al., 2017), indicating the potentially protective effects of H 2 S against mitochondrial toxicity related to ROS (Zhai et al., 2019). Because of the importance of mitochondria in cells, the mechanisms of the regulation by NO and H 2 S on mtDNA oxidative damage in plants under normal or different stresses are intriguing aspects and still need to be deeply studied. However, the current studies about the regulation by NO and H 2 S on mtDNA oxidative damage in plants are still in their infancy, and there is still much research to be done.
CONCLUSION AND PROSPECTS
There are crosstalks among H 2 S, NO, and ROS in the biosynthesis and physiological effects. Both H 2 S and NO can regulate ROS metabolism to maintain redox balance in plants. The redox imbalance causes DNA damage, which in turn exacerbates the imbalance in plants under normal and stress conditions. H 2 S and NO are suggested to protect DNA against damage by indirectly scavenging or removing excessive ROS or by directly modifying the components and improving the ability of the DNA repair pathway. H 2 S, NO, and ROS have many variants and can transform easily and quickly, which brings great difficulties in studying the details and even their inhibitors and scavengers regulating redox balance. At present, a great number of studies focus on the roles of NO in DNA damage repair, but the details and mechanisms that NO regulates damage repair are not clear. Compared with NO, the research on the role of H 2 S in DNA damage repair is in its infancy in plants. The damage and repair pathways of mtDNA are complicated, and the interplays among NO, H 2 S, ROS, and mtDNA damage are also intricate. In what way and with which repair pathways do H 2 S and NO regulate mtDNA oxidative damage in plants under normal or different stresses? Additionally, studies on whether the similar regulation of DNA damage repair by NO and H 2 S present in other organelles, i.e., chloroplasts possessing their DNA (cpDNA) would also be meaningful. In recent years, the development of sequencing technology (including high throughput sequencing and single-cell sequencing) offers a fast and cost-effective method for sequencing the whole mtDNA genome (Grabherr et al., 2011;Sloan, 2013). The exploitation of universal and conserved mitochondrial primers (Duminil, 2014;Pereira et al., 2018), combined with opportunities offered by the availability of complete mtDNA sequence in plant species, facilitate the mtDNA-based molecular studies. Moreover, the emergence of mitochondrial genome editing technology (RNA-free DddAderived cytosine base editors and mitoTALENs) enables the study of mitochondrial gene functions to be carried out in-depth (Kazama et al., 2019;Arimura et al., 2020;Mok et al., 2020). Technological developments may provide the details of mtDNA damage and the roles of NO, H 2 S, and ROS in regulating the repair pathways of mtDNA damage in response to stress in plants as well. More advanced instruments and analytical methods are also needed to study the temporal and spatial changes of NO, H 2 S, and ROS in plants, but there is still a long way to go.
AUTHOR CONTRIBUTIONS
DH and GJ collected the references and completed the first draft. LZ and CC revised the manuscript. SZ designed the framework and edited the manuscript. All authors contributed to the article and approved the submitted version.
FUNDING
This work was supported by the National Natural Science Foundation of China (No. 32071808). | 7,709.8 | 2021-08-06T00:00:00.000 | [
"Medicine",
"Chemistry",
"Environmental Science",
"Biology"
] |
Dominant mixed QCD-electroweak $\mathcal{O}(\alpha_s\alpha)$ corrections to Drell-Yan processes in the resonance region
A precise theoretical description of W- and Z-boson production in the resonance region is essential for the correct interpretation of high-precision measurements of the W-boson mass and the effective weak mixing angle. Currently, the largest unknown fixed-order contribution is given by the mixed QCD-electroweak corrections of $\mathcal{O}(\alpha_s\alpha)$. We argue, using the framework of the pole expansion for the NNLO QCD-electroweak corrections established in a previous paper, that the numerically dominant corrections arise from the combination of large QCD corrections to the production with the large electroweak corrections to the decay of the W/Z boson. We calculate these so-called factorizable corrections of"initial-final"type and estimate the impact on the W-boson mass extraction. We compare our results to simpler approximate combinations of electroweak and QCD corrections in terms of naive products of NLO QCD and electroweak correction factors and using leading-logarithmic approximations for QED final-state radiation as provided by the structure-function approach or QED parton-shower programs. We also compute corrections of"final-final"type, which are given by finite counterterms to the leptonic vectorboson decays and are found to be numerically negligible.
Introduction
The class of Drell-Yan-like processes is one of the most prominent types of particle reactions at hadron colliders and describes the production of a lepton pair through an intermediate gaugeboson decay, pp/pp → V → 1¯ 2 + X.
Depending on the electric charge of the colour-neutral gauge boson V , the process can be further classified into the neutral-current (V = Z/γ) and the charged-current (V = W ± ) processes. The large production rate in combination with the clean experimental signature of the leptonic vector-boson decay allows this process to be measured with great precision. Moreover, the Dell-Yan-like production of W or Z bosons is one of the theoretically best understood and most precisely predicted processes. As a consequence, electroweak (EW) gauge-boson production is among the most important "standard-candle" processes at the LHC (see, e.g. Refs. [1,2]). Its cross section can be used as a luminosity monitor, and the measurement of the mass and width of the Z boson represents a powerful tool for detector calibration. Furthermore, the W charge asymmetry and the rapidity distribution of the Z boson deliver important constraints in the fit of the parton distribution functions (PDFs) [3], which represent crucial ingredients for almost all predictions at the LHC. Of particular relevance for precision tests of the Standard Model is the potential of the Drell-Yan process at the LHC for high-precision measurements in the resonance regions, where the effective weak mixing angle, quantified by sin 2 θ eff , might be extracted from data with LEP precision [4]. The W-boson mass can be determined from a fit to the distributions of the lepton transverse momentum (p T, ) and the transverse mass of the lepton pair (M T, ν ) which exhibit Jacobian peaks around M W and M W /2, respectively, and allow for a precise extraction of the mass with a sensitivity below 10 MeV [5,6] provided that PDF uncertainties can be reduced [7][8][9][10].
In addition to the N 3 LO QCD corrections, the next frontier in theoretical fixed-order computations is given by the calculation of the mixed QCD-EW corrections of O(α s α) [53]. These corrections can affect observables relevant for the M W determination at the percent level [54] and therefore must be under theoretical control. Up to now, QCD and EW corrections have been combined in various approximations [55][56][57][58][59][60]. However, a full NNLO calculation at O(α s α) is necessary for a proper combination of QCD and EW corrections without ambiguities. Here some partial results for two-loop amplitudes [61][62][63] as well as the full O(α s α) corrections to the W and Z decay widths [64,65] are known. A complete calculation of the O(α s α) corrections requires to combine the double-virtual corrections with the O(α) EW corrections to W/Z + jet production [66][67][68][69][70][71][72], the O(α s ) QCD corrections to W/Z + γ production [68,[73][74][75][76][77][78][79][80][81], and the double-real corrections using a method to regularize infrared (IR) singularities.
In a previous paper [82], we have initiated the calculation of the O(α s α) corrections to Drell-Yan processes in the resonance region via the so-called pole approximation (PA) [83], which has been successfully applied to the EW corrections to W production [39,82,84,85] and Z production [82] at NLO. It is based on a systematic expansion of the cross section about the resonance pole and is suitable for theoretical predictions in the vicinity of the gauge-boson resonance, where the higher precision is especially relevant. The PA splits the corrections into distinct well-defined subsets, which can be calculated separately. This allows to assess the numerical impact of different classes of corrections and to identify the dominant contributions. More precisely, the contributions can be classified into two types: the factorizable and the non-factorizable corrections. In the former, the corrections can be separately attributed to the production and the subsequent decay of the gauge boson, whereas in the latter the production and decay subprocesses are linked by the exchange of soft photons. At O(α), the PA shows agreement with the known NLO EW corrections up to fractions of 1% near the resonance, i.e. at a phenomenologically satisfactory level [82]. In particular, the bulk of the NLO EW corrections near the resonance can be attributed to the factorizable corrections to the W/Z decay subprocesses, while the factorizable corrections to the production process are mostly suppressed below the percent level, and the non-factorizable contributions being even smaller.
Based on the quality of the PA at NLO we are confident that this approach is suitable to calculate the O(α s α) corrections with sufficient accuracy for the description of observables that are dominated by the resonances. The non-factorizable corrections comprise the conceptually most challenging contribution to the PA and have been computed at O(α s α) in Ref. [82]. They turn out to be very small and, thus, demonstrate that for phenomenological purposes the O(α s α) corrections can be factorized into terms associated with initial-state and/or final-state corrections and combinations of the two types. In this paper we calculate the factorizable corrections of the type "initial-final", which combine large QCD corrections to the production with the large EW corrections to the decay of the W/Z boson. Therefore we expect to capture the dominant contribution at O(α s α) to observables relevant for precision physics dominated by the W and Z resonances. We also compute the corrections of "final-final" type, which are given only by finite counterterms to the leptonic vector-boson decay. The remaining factorizable "initial-initial" corrections are expected to deliver only a small contribution and would further require O(α s α)corrected PDFs for a consistent evaluation, which are however not available. It is all the more important to isolate this contribution in a well-defined manner, as it is accomplished by the PA.
A technical aspect of higher-order calculations involving massless particles is the proper treatment of IR singularities that are associated with configurations involving soft and/or collinear particles. To this end, we use the dipole subtraction formalism [86][87][88][89] and its extension for decay processes presented in Ref. [90] for the analytic cancellation of all IR singularities. Although the cancellation of IR singularities in the O(α s α) corrections presented in this work is accomplished by using a combined approach of the techniques developed for NLO calculations, it represents one of the main technical difficulties in the calculation and we devote special attention to its discussion.
This paper is organized as follows: In Section 2 we present the calculation of the initial-final and final-final factorizable corrections. We discuss the construction of an IR-finite final result for the initial-final corrections in detail with a special focus on the treatment of the combined IR singularities of the QCD and EW corrections. Our numerical results are presented in Section 3, where we compare them to different versions of a naive product ansatz obtained by multiplying NLO QCD and EW correction factors, and to a leading-logarithmic treatment of photon radiation as provided by the structure-function approach or QED parton showers such as PHOTOS [91]. We further perform a χ 2 fit in order to estimate the effect of the NNLO O(α s α) corrections on the measurement of the W-boson mass. A summary is given in Sect. 4.
Calculation of the dominant O(α s α) corrections in pole approximation
In this section we identify and calculate the dominant O(α s α) corrections to the chargedcurrent and neutral-current Drell-Yan processes in the vicinity of an intermediate vector-boson resonance. In Sect. 2.1 we describe the classification of the O(α s α) corrections in the framework of the PA [82]. We identify factorizable contributions of "initial-final" type-i.e. the combination of QCD corrections to vector-boson production with EW corrections to vector-boson decayas dominant source for corrections to distributions dominated by the vector-boson resonance. The calculation of the building blocks contributing to the initial-final factorizable corrections is performed in Sect. 2.2. In Sect. 2.3 the different building blocks of the initial-final contributions are combined into a formula suitable for numerical evaluation, where all IR singularities are cancelled explicitly. Finally, in Sect. 2.4 we calculate corrections of "final-final" type, which are given by pure counterterm contributions and are numerically small.
Survey of types of O(α s α) corrections in pole approximation
The PA for Drell-Yan processes [39,[82][83][84][85] provides a systematic classification of contributions to Feynman diagrams that are enhanced by the resonant propagator of a vector boson V = W, Z. The leading corrections in the expansion around the resonance pole arise from factorizable corrections to W/Z production and decay subprocesses, and non-factorizable corrections that link production and decay by soft-photon exchange. The PA separates corrections to production and decay stages in a consistent and gauge-invariant way. This is particularly relevant for the charged-current Drell-Yan process, where photon radiation off the intermediate W boson contributes simultaneously to the corrections to production and decay of a W boson, and to the non-factorizable contributions. Applications of different variants of the PA to NLO EW corrections [39,82,84,85] have been validated by a comparison to the complete EW NLO calculations and show excellent agreement at the order of some 0.1% in kinematic distributions dominated by the resonance region. The structure of the PA for the O(α s α) correction has been worked out in Ref. [82], where details of the method and our setup can be found. The corrections can be classified into the four types of contributions shown in Fig. 1 for the case of the double-virtual corrections. For each class of contributions with the exception of the final-final corrections (c), also the associated real-virtual and double-real corrections have to be computed, obtained by replacing one or both of the labels α and α s in the blobs in Fig. 1 by a real photon or gluon, respectively. The corresponding crossed partonic channels, e.g. with quark-gluon initial states have to be included in addition.
In detail, the four types of corrections are characterized as follows: (a) The initial-initial factorizable corrections are given by two-loop O(α s α) corrections to onshell W/Z production and the corresponding one-loop real-virtual and tree-level double-real contributions, i.e. W/Z + jet production at O(α), W/Z + γ production at O(α s ), and the processes W/Z + γ + jet at tree level. Results for individual ingredients of the initialinitial part are known, such as partial two-loop contributions [61,63] and the full O(α) EW corrections to W/Z+jet production including the W/Z decays [69][70][71]. However, a consistent combination of these building blocks requires also a subtraction scheme for IR singularities at O(α s α) and has not been performed yet. Note that currently no PDF set including O(α s α) corrections is available, which is required to absorb IR singularities of the initial-initial corrections from photon radiation collinear to the beams.
Results of the PA at O(α) show that observables such as the transverse-mass distribution in the case of W production or the lepton-invariant-mass distributions for Z production are extremely insensitive to initial-state photon radiation [82]. Since these distributions also receive relatively moderate QCD corrections, we do not expect significant initial-initial NNLO O(α s α) corrections to such distributions. For observables sensitive to initial-state recoil effects, such as the transverse-lepton-momentum distribution, the O(α s α) corrections should be larger, but still very small compared to the huge QCD corrections. 1 (b) The factorizable initial-final corrections consist of the O(α s ) corrections to W/Z production combined with the O(α) corrections to the leptonic W/Z decay. Both types of corrections are large and have a sizable impact on the shape of differential distributions at NLO, so that we expect this class of the factorizable corrections to capture the dominant O(α s α) effects. The computation of these contributions is the main result of this paper and is discussed in Sect. 2.2. Preliminary numerical results of these corrections were presented in Refs. [92,93].
(c) Factorizable final-final corrections arise from the O(α s α) counterterms of the lepton-W/Zvertices, which involve only QCD corrections to the vector-boson self-energies. There are no corresponding real contributions, so that the final-final corrections have practically no impact on the shape of distributions. We compute these corrections in Sect. 2.4 below and confirm the expectation that they are phenomenologically negligible.
(d) The non-factorizable O(α s α) corrections are given by soft-photon corrections connecting the initial state, the intermediate vector boson, and the final-state leptons, combined with QCD corrections to V -boson production. As shown in detail in Ref. [82], these corrections into initial-initial (b) and initial-final (c) parts, illustrated for an example in part (a). The momentum p V of the intermediate vector boson V is given by can be expressed in terms of soft-photon correction factors to squared tree-level or oneloop QCD matrix elements by using gauge-invariance arguments. The numerical impact of these corrections was found to be below the 0.1% level and is therefore negligible for all phenomenological purposes.
The definition of the factorizable corrections and the separation of initial-and final-state corrections is illustrated in Fig. 2 for the case of the double-real corrections. An example diagram for the charged-current process is given in Fig. 2a, which cannot be attributed uniquely to the vector-boson production or decay subprocess and displays an overlapping resonance structure due to the propagator poles at p 2 V = µ 2 V and (p V + k) 2 = µ 2 V . Here µ V combines the real mass and width parameters of V , M V and Γ V , to a complex mass value, However, a simple partial-fractioning identity for the two V -boson propagators allows us to disentangle the two resonance structures and to decompose such diagrams into contributions associated with photon emission from the production or decay subprocesses of an on-shell V boson (see Eq. (2.11) in Ref. [82]). This is illustrated in Fig. 2a, where the double slash on a propagator line indicates that the corresponding momentum is set on its mass shell in the rest of the diagram (but not on the slashed line itself). Using this decomposition, the double-real corrections can be divided consistently into initial-initial and initial-final contributions, as shown in Fig. 2b and Fig. 2c, respectively. Here a diagrammatic notation is used where an encircled diagram with an attached photon or gluon stands for all possibilities to attach the photon/gluon to the fermion line and the gauge boson V (see Eq. (2.12) in Ref. [82] for an example). The initial-final (virtual QCD)×(real EW) corrections are treated analogously. All different contributions to the factorizable initialfinal corrections are diagrammatically characterized in terms of interference diagrams in Fig. 3.
Calculation of the factorizable initial-final corrections
In this section we calculate the various contributions to the factorizable initial-final corrections of O(α s α) shown in Fig. 3. Most contributions can be expressed in terms of reducible products of NLO QCD and NLO EW building blocks. For details on the notation used for these NLO results we refer to Ref. [82].
Double-virtual corrections
The double-virtual O(α s α) initial-final corrections to the squaredq a q b → 1¯ 2 amplitude are illustrated in Fig. 3(a) in terms of interference diagrams. They arise in two ways: from the interference of the tree amplitude with the two-loop O(α s α) amplitude and from the interference between the one-loop amplitudes with O(α s ) corrections to V -boson production and O(α) corrections to the decay, respectively, Vs,PA * . (2.1) The LO amplitude in PA, M 0,PA , differs from the full LO matrix element by the absence of the non-resonant photon diagram in case of the neutral-current Drell-Yan process. The first term on the right-hand side in Eq. (2.1) involves the factorizable initial-final contribution to the two-loop amplitude, which takes the form of reducible (one-loop)×(one-loop) diagrams and is defined explicitly as where a sum over the physical polarization states of the vector boson V , labelled by λ V , is performed. In the second step in Eq. (2.2) the fact is used that the one-loop corrections to the production and decay factorize off the corresponding LO matrix elements, The virtual QCD corrections are well known and are quoted explicitly in Eq. (2.35) of Ref. [82]. The explicit expressions for the NLO EW correction factors can be found, e.g., in Refs. [39,47], and are quoted in Appendix B.2 of Ref. [93]. In order to maintain gauge invariance, the NLO production and decay subamplitudes in Eq. (2.2), and in particular the correction factor δ dec Vew , are evaluated for on-shell V bosons. We keep the QCD correction factor δ Vqaq b Vs off shell, i.e. without setting s → M 2 V there to be closer to the full calculation, which is possible, because δ Vqaq b Vs does not depend on M V at all. The on-shell projection s → M 2 V in the EW correction involves some freedom, but numerical effects from different implementations are of the same order as the intrinsic uncertainty of the PA. However, the choice of the mappings in the virtual and real corrections has to match properly in order to ensure the correct cancellation of IR singularities. Both the EW and QCD correction factors contain soft and collinear singularities, which take the form of 1 2 poles for massless fermions. Therefore, in principle, Eq. (2.5) requires the evaluation of the correction factors up to O 2 in order to obtain all finite O 0 terms. However, after applying the subtraction formalism, which we describe in detail in Sect. 2.3, the poles are cancelled before performing the full expansion in and, thus, the results up to order O 0 turn out to be sufficient. This result is obvious if the soft and collinear singularities are not regularized in D = 4 − 2 dimensions, but by small mass parameters, where no rational terms from the multiplication of 1/ poles with D-dimensional quantities exist at all.
(Real QCD)×(virtual EW) corrections
The (real QCD)×(virtual EW) contributions to the factorizable initial-final corrections shown in Fig. 3(b) arise by including the virtual corrections to the leptonic W/Z decays to the various partonic subprocesses of V + jet production, For the quark-induced channel, the corrections are given by replacing the virtual QCD amplitude in Eq. (2.2) by the corresponding amplitude for real-gluon emission, Analogously to the double-virtual case, the EW decay subamplitude is evaluated for on-shell vector bosons, while the QCD correction is kept off shell. Using the factorization property of the EW one-loop decay corrections (2.4), the (real QCD)×(virtual EW) correction to the cross section in the quark-anti-quark channel is proportional to the real NLO QCD corrections dσ Rs , As for the Born amplitude, the label PA in the real-emission corrections indicates that all nonresonant terms, i.e. the photon-exchange diagrams in case of the neutral-current process, are omitted in the QCD real-emission amplitudes. Analogous expressions hold for the gluon-quark and gluon-anti-quark initiated subprocesses to V + jet production.
(Virtual QCD)×(real photonic) corrections
The (virtual QCD)×(real photonic) factorizable corrections of initial-final type arise from the generic interference diagram shown in Fig. 3(c). They are obtained by combining the real-photon corrections to on-shell V -boson decay with the virtual QCD corrections to V -boson production, In the second step, Eq. (2.3) has been used to factorize the virtual QCD correction factor from the matrix element Mq aqb→ 1¯ 2 γ Rew,fact,dec for the factorizable NLO decay corrections (see Eq. (2.14) in Ref [82]). Again the matrix elements for the EW decay subprocess is evaluated for on-shell vector bosons, while the QCD correction factor is kept off shell. As illustrated in Fig. 2 and discussed in detail in Ref. [82], the splitting of photon-emission effects off the intermediate V -boson into parts corresponding to initial-or final-state radiation separates the two resonance propagator For factorizable EW decay correction we, thus, have to perform the on-shell projection (p 2 The resulting contribution of the (virtual QCD)×(real photonic) corrections to the cross section therefore assumes the form (2.10)
Double-real corrections
The double-real emission corrections are illustrated by interference diagrams in Fig. 3(d) and are defined by the real-emission matrix elements for the V + jet production subprocesses (2.6) with the subsequent decay V → 1¯ 2 γ, with analogous expressions for the gq and gq channels. The non-resonant contribution arising from the case V = γ in the neutral-current process is again not included. Compact explicit results for the helicity amplitudes of the double-real corrections can be found in Ref. [93]. The double-real contribution to the cross section, dσ Rs⊗Rew prod×dec , is defined in terms of the square of the matrix element (2.11) where the decay subamplitudes are evaluated for on-shell V bosons. Due to the spin correlations of the production and decay matrix elements and the full kinematics of the 2 → 4 scattering process, the double-real corrections do not factorize further into separate EW and QCD correction factors, in contrast to the other classes of factorizable initial-final corrections.
Treatment of infrared singularities for the factorizable initial-final corrections
The NNLO O(α s α) contributions to the cross section due to the factorizable initial-final corrections are obtained by integrating the four contributions discussed in the previous section over the respective phase spaces, where the additional QCD collinear counterterms in the last line were introduced to absorb the collinear singularities associated with the quarks and gluons in the initial state into the NLO PDFs. Note that the EW corrections are completely confined to the decay subprocess, and consequently, there are no singularities from initial-state collinear quark-photon splittings. This allows us to obtain the collinear counterterms in the last line of Eq. Applying the QCD dipole subtraction formalism [86] in order to cancel the IR singularities associated with the QCD corrections, Eq. (2.13) can be written in the following form, (2.14) The explicit expressions of the dipole operators dV dip and the insertion operators I, K, and P can be found in Ref. [86]. The symbol ⊗ denotes possible additional helicity and colour correlations, and it is implicitly assumed that the cross sections multiplying the dipole operators dV dip are evaluated on the respective dipole-mapped phase-space point. The explicit expressions associated with the NLO QCD corrections were given in Ref. [82]. All individual integrals appearing in Eq. (2.14) are now free of QCD singularities, but remain IR divergent owing to the singularities contained in the EW corrections which still need to be cancelled between the virtual corrections and the corresponding real-photon-emission parts. For this purpose we employ the dipole subtraction formalism for photon radiation [87,89], in particular the extension of the formalism to treat decay kinematics described in detail in Ref. [90]. As a result, we are able to arrange the six contributions in Eq. (2.14) into a form where all IR divergences are cancelled in the integrands explicitly, where each term is an IR-finite object and its phase-space integration can be performed numerically in four dimensions. Equation (2.15) is our master formula for the numerical evaluation discussed in Sect. 3. Explicit expressions for all contributions for the quark-anti-quark and quark-gluon induced channels are given in Appendix B. The first two terms in Eq. (2.15) arise from the sum of the double-virtual and the (virtual QCD)×(real photonic) corrections, including the insertion operators from the QCD dipole formalism, and correspond to the sum of the first two lines in Eq. (2.14). Applying the dipole formalism to rearrange the IR singularities of photonic origin between the virtual and real EW corrections, we obtain the following expressions for the IR-finite virtual QCD contributions to the cross section, where the sum over the emitter-spectator pairs (I, J) in Eq. (2.17) extends over all particles of the decay subprocess, i.e. I, J = 1 ,¯ 2 , V . We have introduced a compact notation for the QED dipoles, where η i = 1 for incoming particles and outgoing antiparticles and η i = −1 for incoming antiparticles and outgoing particles. The corresponding endpoint contributions are given by where the functions g (sub) and G (sub) are given in Ref. [87], while d (sub) and D (sub) are the decay dipoles and their integrated counterparts constructed in Ref. [90]. Whenever we write 1 ↔¯ 2 , this implies the interchange Q 1 ↔ Q 2 of the electric charges of the respective fermions, irrespective of their particle or antiparticle nature.
As anticipated in Sect. 2.2.1, all IR singularities contained in δ Vqaq b Vs cancel exactly against the corresponding poles of the I operator within the second square bracket of Eq. (2.16). Similarly, all singularities in δ dec Vew cancel against the corresponding poles in I ew in the first square bracket of Eq. (2.16). As a consequence, it is sufficient to use the correction factors δ dec Vew and δ Vqaq b Vs up to O 0 . Furthermore, we recall that the correction factors δ dec Vew are evaluated at the on-shell point p 2 V = M 2 V and, thus, are independent of the phase-space kinematics. The contributions involving real QCD corrections are given by the third and forth term in Eq. (2.15). They are obtained by applying the QED dipole subtraction formalism to the sum of the third and forth line of Eq. (2.14) and result in the following expressions for the IR-finite real-gluon contributions to the cross section, It is instructive to examine the local cancellation of the IR singularities in Eq. (2.21) in more detail. The second term inside the curly brackets of Eq. (2.21) acts as a local counterterm to the double-real emission cross section dσ Rs⊗Rew in all regions of phase space where the additional QCD radiation becomes unresolved, i.e. soft and/or collinear to the beam. The third term inside the curly brackets of Eq. (2.21) analogously ensures the cancellation of IR singularities in the phase-space regions where the photon becomes soft and/or collinear to a final-state lepton. A subtlety arises in the double-unresolved cases, where the cross sections dσ Rew dec and dσ Rs PA become singular as well, and both subtraction terms above will simultaneously act as a local counterterm, leading to the twofold subtraction of the IR singularities. This disparity in the double-unresolved limits is exactly compensated by the last term inside the curly brackets of Eq. (2.21), which therefore has the opposite sign. Note that the evaluation of this last term involves the successive application of two dipole phase-space mappings. Owing to the property of the factorizable initial-final corrections where the emissions in the production and decay stages of the V boson proceed independently, the two dipole mappings do not interfere with each other and the order in which they are applied is irrelevant. A related property is the factorization of the dipole phase space, where the two one-particle subspaces associated with the two unresolved emissions can be isolated simultaneously. This has the important consequence that the analytic integration over the gluon and photon momenta can be carried out in the same manner as at NLO, which allows us to reuse the known results for the integrated dipoles without modification.
Finally, we consider the convolution terms with additional virtual or real EW corrections given by the last two terms in Eq. (2.15). Since these contributions are essentially given by the lower-order (in α s ) cross sections, convoluted with the insertion operators K and P, they pose no additional complications, and the resulting IR-finite contributions to the cross section can be written as Owing to the Lorentz invariance of the dipole formalism, no special treatment is required in contrast to our calculation of the non-factorizable corrections discussed in Ref. [82], which was carried out with the slicing method to isolate soft-photon singularities.
The results presented so far are appropriate for the case of IR-safe observables, i.e. for the case where collinear photons and leptons are recombined to a "dressed" lepton carrying their total momentum. For non-collinear-safe observables with respect to the final-state leptons, i.e. the treatment of bare muons without photon recombination, we use the method of Ref. [89] and its extension to decay kinematics described in Ref. [90]. The required modifications are described in Appendix C.
Factorizable final-final corrections
The factorizable NNLO corrections of final-final type arise purely from the counterterms to the V 1¯ 2 vertex and therefore factorize from the LO matrix element, The counterterms for the leptonic vector-boson decay only receive contributions from the vectorboson self-energies at O(α s α) [94][95][96][97][98][99], which enter the counterterms through the vector-boson wave-function renormalization constants and through the renormalization constants of the electromagnetic coupling and the weak-mixing angle. There is only one type of contribution from one-loop diagrams with insertions of one-loop O(α s ) or O(α) counterterms. It results from massive quark loops in the vector-boson self-energies where the QCD mass renormalization constant has to be taken into account. We make use of the expressions for the vector-boson selfenergies of Ref. [99], which include the QCD quark mass counterterm in the on-shell scheme.
The expressions for the self-energies in terms of the scalar integrals computed in Ref. [99] are given in Appendix A.
The vertex counterterms in the on-shell renormalization scheme are obtained from the expressions for the corresponding NLO EW counterterms [100] upon replacing the one-loop vectorboson self-energies by the two-loop O(α s α) results and dropping lepton wave-function renormalization constants, which receive no correction at this order. We employ the G µ input-parameter scheme where the electromagnetic coupling constant is derived from the Fermi constant G µ via the relation The counterterm δZ e for the electromagnetic charge in the G µ scheme is related to the one in the α(0) input-parameter scheme as follows, The quantity ∆r comprises all higher-order corrections to muon decay excluding the contributions that constitute QED corrections in the Fermi model, which are included in the definition of the muon decay constant G µ [101], The final-final correction to the cross section for the charged-current cross section is therefore below the 0.1% level and phenomenologically negligible. This can be partially attributed to the choice of the G µ -scheme where universal corrections to charged-current processes are absorbed in the value of α Gµ . The numerical values of the counterterms δ ct,τ,(αsα) Z ¯ for the Z ¯ vertices with lepton chiralities τ = ± are somewhat larger, but of opposite sign: The resulting corrections to the neutral-current Drell-Yan process are, however, suppressed far below the 0.1% level due to cancellations between the right-and left-handed production channels and are therefore also negligible for all phenomenological purposes.
Numerical results
In this section we present the numerical results for the dominant mixed QCD-EW corrections to the Drell-Yan process at the LHC for a centre-of-mass energy of √ s = 14 TeV. We consider the two processes with electrons or muons in the final state ( = e, µ). We further distinguish two alternative treatments of photon radiation: In the "dressed-lepton" case, collinear photon-lepton configurations are treated inclusively using a photon-recombination procedure. As a result, the numerical predictions do not contain large logarithms of the lepton mass, which can be set to zero. The dressed-lepton results are appropriate mostly for electrons in the final state. In the "bare-muon" case, no such recombination is performed, reflecting the experimental situation which allows for the detection of isolated muons. We perform a comparison to naive factorization prescriptions of QCD and EW corrections, as well as to a modelling of photonic final-state radiation (FSR) by structure functions or a photon shower. Moreover, we estimate the impact of the NNLO QCD-EW corrections on the measurement of the W-boson mass.
Input parameters and event selection
The setup for the calculation is analogous to the one used in Ref. [82]. The choice of input parameters closely follows Ref. [102], We convert the on-shell masses and decay widths of the vector bosons to the corresponding pole masses and widths as spelled out in Ref. [82]. The electromagnetic coupling constant used in the LO predictions is obtained from the Fermi constant by Eq. (2.25). In the charged-current process, all relative electroweak corrections are computed using α Gµ . In the neutral-current process, however, we follow Ref. [47] and use α(0) consistently in the relative photonic corrections while the remaining relative weak corrections are proportional to α Gµ . The same prescription is applied to the relative O(α s α) corrections.
The masses of the light quark flavours (u, d, c, s, b) and of the leptons are neglected throughout, with the only exception in case of non-collinear-safe observables, where the final-state collinear singularity is regularized by the mass of the muon, m µ = 105.658369 MeV. (3. 2) The CKM matrix is chosen diagonal in the third generation and the mixing between the first two generations is parametrized by the following values for the entries of the quark-mixing matrix, For the PDFs we consistently use the NNPDF2.3 sets [103], where the NLO and NNLO QCD-EW corrections are evaluated using the NNPDF2.3QED NLO set [104], which also includes O(α) corrections. The value of the strong coupling α s (M Z ) quoted in Eq. (3.1) is dictated by the choice of these PDF sets. For the evaluation of the full NLO EW corrections entering the naive products below, we employ the DIS factorization scheme to absorb the mass singularities into the PDFs. The renormalization and factorization scales are set equal, with a fixed value given by the respective gauge-boson mass, in order to avoid the photon pole at M → 0. For the dressed-lepton case, in addition, a photon recombination procedure analogous to the one used in Refs. [39,47] is applied: 1. Photons close to the beam with a rapidity |η γ | > 3 are treated as beam remnants and are not further considered in the event selection.
2. For the photons that pass the first step, the angular distance to the charged leptons is computed, where φ denotes the azimuthal angle in the transverse plane. If the distance R ± γ between the photon and the closest lepton is smaller than 0.1, the photon is recombined with the lepton by adding the respective four-momenta, ± (k i ) + γ(k) → ± (k i + k).
Results for the dominant factorizable corrections
The NNLO QCD-EW corrections to the hadronic Drell-Yan cross section are dominated by the factorizable initial-final O(α s α) corrections, ∆σ NNLO s⊗ew prod×dec , which are obtained by convoluting the corresponding partonic correctionsσ NNLO s⊗ew prod×dec calculated in Section 2 with the PDFs. Our default prediction for Drell-Yan processes is then obtained by adding these NNLO corrections to the sum of the full NLO QCD and EW corrections, where all terms are consistently evaluated with the NNPDF2.3QED NLO PDFs. The nonfactorizable corrections computed in Ref. [82] were found to have a negligible impact on the cross section and are therefore not included here. Similarly, the factorizable corrections of "final-final" type discussed in Sect. 2.4 turn out to have a negligible impact on the cross-section prediction and are therefore not included in Eq. (3.8) either. Our result allows to validate estimates of the NNLO QCD-EW corrections based on a naive product ansatz. For this purpose, we define the naive product of the NLO QCD cross section and the relative EW corrections, where the relative EW corrections are defined as the ratio of the NLO EW contribution ∆σ NLOew with respect to the LO contribution σ 0 according to we can cast the relative difference of our best prediction (3.8) and the product ansatz (3.9) into the following form, where the LO prediction σ LO in the denominators is evaluated with the LO PDFs. The difference of the relative NNLO correction δ prod×dec αsα and the naive product δ αs δ (dec) α therefore allows to assess the validity of a naive product ansatz. As observed in Sect. 2.2, most contributions to the factorizable initial-final corrections take the reducible form of a product of two NLO 2 Note that the correction factor δ αs differs from that in the standard QCD K factor KNLO s = σNLO s /σLO ≡ 1 + δα s due to the use of different PDF sets in the Born contributions. See Ref. [92] for further discussion. can therefore be attributed to this type of contribution. The difference of the naive product defined in terms of δ dec α and δ α allows us to assess the impact of the missing O(α s α) corrections beyond the initial-final corrections considered in our calculation and therefore also provides an error estimate of the PA, and in particular of the omission of the corrections of initial-initial type. Figure 4 shows the numerical results for the relative O(α s α) initial-final factorizable corrections δ prod×dec αsα to the transverse-mass (M T,ν ) and the transverse-lepton-momentum (p T, ) distributions for W + production at the LHC. For Z production, Figure 5 Eq. (3.12). In the following, we mainly focus on the results for bare muons. The respective results with photon recombination display the same general features as those for bare muons, but the relative corrections are reduced by approximately a factor of two. This reduction is familiar from NLO EW results and is induced by the cancellation of the collinear singularities by restoring the level of inclusiveness required for the KLN theorem. One observes that the NNLO δ prod×dec αsα corrections are in general better approximated by the simple product ansatz for the case of bare muons than for dressed leptons. This can be understood from the fact that the dominant part of the corrections stem from the collinear logarithms ln(m µ ) which are known to factorize.
For the M T,ν distribution for W + production (upper left plot in Fig. 4), the mixed NNLO QCD-EW corrections for bare muons are moderate and amount to approximately −1.7 % around the resonance, which is about an order of magnitude smaller than the NLO EW corrections. Both variants of the naive product provide a good approximation to the full result in the region around and below the Jacobian peak, which is dominated by resonant W production. For larger values of M T,ν , the product δ αs δ α based on the full NLO EW correction factor deviates from the other curves, which signals the growing importance of effects beyond the PA. However, the deviations amount to only few per-mille for M T,ν 90 GeV. The overall good agreement between the δ prod×dec αsα corrections and both naive products can be attributed to well-known insensitivity of the observable M T,ν to initial-state radiation effects already seen in the case of NLO corrections in Ref [82].
For the p T, distributions in the case of bare muons (upper right plots in Figs. 4 and 5, respectively) we observe corrections that are small far below the Jacobian peak, but which rise to about 15% (20%) on the Jacobian peak at p T, ≈ M V /2 for the case of the W + boson (Z boson) and then display a steep drop reaching almost −50% at p T, = 50 GeV. This enhancement stems from the large QCD corrections above the Jacobian peak familiar from the NLO QCD results (see e.g. Fig. 8 in Ref. [82]) where the recoil due to real QCD radiation shifts events with resonant W/Z bosons above the Jacobian peak. The naive product ansatz fails to provide a good description of the full result δ prod×dec αsα and deviates by 5-10% at the Jacobian peak, where the PA is expected to be the most accurate. This can be attributed to the strong influence of the recoil induced by initial-state radiation on the transverse momentum, which implies a larger effect of the double-real emission corrections on this distribution that are not captured correctly by the naive products. The two versions of the naive products display larger deviations than in the M T,ν distribution discussed above, which signals a larger impact of the missing O(α s α) initialinitial corrections. However, these deviations should be interpreted with care, since a fixed-order prediction is not sufficient to describe this distribution around the peak region p T, ≈ M V /2, which corresponds to the kinematic onset for V + jet production and is known to require QCD resummation for a proper description.
In case of the M distribution for Z production (left-hand plots in Fig. 5), corrections up to 10% are observed below the resonance for the case of bare muons. This is consistent with the large EW corrections at NLO in this region, which arise from final-state photon radiation that shifts the reconstructed value of the invariant lepton-pair mass away from the resonance to lower values. The naive product approximates the full initial-final corrections δ prod×dec αsα reasonably well at the resonance itself (M = M Z ) and above, but completely fails already a little below the resonance where the naive products do not even reproduce the sign of the full δ prod×dec αsα correction. This deviation occurs although the invariant-mass distribution is widely unaffected by initial-state radiation effects. The fact that we obtain almost identical corrections from the two versions of the product δ αs δ dec α and δ αs δ α demonstrates the insensitivity of this observable to photonic initial-state radiation.
In order to locate the source of this large discrepancy we examine the individual correction factors in the naive product in more detail. We restrict ourselves to the case of bare muons and the full NLO EW correction factor δ α for definiteness, which does not affect our conclusions. In Fig. 6(a), we separately plot the two correction factors that enter the naive product δ αs × δ α and further divide the QCD corrections into the qq-and the qg-induced contributions. We observe that the two different qq-and qg-induced channels individually receive large QCD corrections, however, they differ in sign, so that large cancellations take place in the sum δ αs . A small mismatch in the corrections of the individual channels can therefore quickly lead to a large effect in the QCD corrections which is then further enhanced by the large EW corrections in the product ansatz δ αs × δ α . Moreover, Figure 6(a) reveals that the QCD correction factor δ αs is responsible for the sign change at M ≈ 83 GeV which is the most striking disagreement of the naive product ansatz with the full factorizable initial-final corrections. This zero crossing happens more than three widths below the resonance where the cross section is reduced by almost two orders of magnitudes compared to the resonance region and, furthermore, will be very sensitive to event selection cuts, since δ αs arises from the cancellation of two large corrections as we have seen above. In Fig. 6(a) we observe the large EW corrections below the resonance mentioned above, which arise due to the redistribution of events near the Z pole to lower lepton invariant masses by final-state photon radiation. The similar form of the factorizable NNLO initial-final corrections indicates that they mainly stem from an analogous mechanism. This suggests that it is more appropriate to replace the QCD correction factor δ αs in the naive product by its value at the resonance δ αs (M = M Z ) ≈ 6.5%, which corresponds to the location of the events that are responsible for the bulk of the large EW corrections below the resonance. In contrast, the naive product ansatz simply multiplies the corrections locally on a bin-by-bin basis. This causes a mismatch in the correction factors and fails to account for the migration of events due to FSR. The comparison of the previous results and the modified product is shown in Fig. 6(b) and clearly shows an improvement despite its very crude construction. Contrary to the lepton-invariant-mass distribution, the transverse-mass distribution is dominated by events with resonant W bosons even in the range below the Jacobian peak, M T,ν M W , so it is less sensitive to the redistribution of events to lower M T,ν . This explains why the naive product can provide a good approximation of the full initial-final NNLO corrections. It should be emphasized, however, that even in the case of the M T,ν distribution any event selection criteria that deplete events with resonant W bosons below the Jacobian peak will result in increased sensitivity to the effects of FSR and can potentially lead to a failure of a naive product ansatz.
In conclusion, simple approximations in terms of products of correction factors have to be used with care and require a careful case-by-case investigation of their validity.
Leading-logarithmic approximation for final-state photon radiation
As is evident from Figs. 4 and 5, a naive product of QCD and EW correction factors (3.9) is not adequate to approximate the NNLO QCD-EW corrections for all observables. A promising approach to a factorized approximation for the dominant initial-final corrections can be obtained by combining the full NLO QCD corrections to vector-boson production with the leadinglogarithmic (LL) approximation for the final-state corrections. The benefit in this approximation lies in the fact that the interplay of the recoil effects from jet and photon emission is properly taken into account. On the other hand, the logarithmic approximation neglects certain (nonuniversal) finite contributions, which are, however, suppressed with respect to the dominating radiation effects.
In the structure-function approach [105], the leading-logarithmic approximation of the photonic decay corrections is combined with the NLO QCD corrections to the production by a convolution, where dσ NLOs includes the virtual and real QCD corrections. The step function Θ cut (z i k i ) is equal to 1 if the event passes the cut on the rescaled lepton momentum z i k i and 0 otherwise. The variables z i are the momentum fractions describing the respective lepton energy loss by collinear photon emission. For the charged-current process only one of the convolutions is present. The O(α) contribution to the structure function Γ LL reads where the large mass logarithm appears in the variable and Q denotes the relative electric charge of the lepton . In order to be consistent in the comparison with our calculation as described in Sect. 3.1, the electromagnetic coupling constant α appearing in Eq. (3.15) is set to α Gµ and α(0) for the charged-current and neutral-current processes, respectively. The scale Q is chosen as the gauge-boson mass, Since the mass logarithms cancel in observables where photon emission collinear to the final-state charged leptons is treated fully inclusively, the structure-function approach is only applicable to non-collinear-safe observables, i.e. to the bare-muon case. initial-state QCD and final-state EW corrections to our best prediction δ prod×dec αsα for the case of the transverse-mass (left) and transverse-lepton-momentum (right) distributions for W + production at the LHC, as in Fig. 4. In the bare-muon case, the result (3.13) of the structure-function approach is also shown.
In contrast, in parton-shower approaches to photon radiation (see e.g. Refs. [49,50,106]) the photon momenta transverse to the lepton momentum are generated as well, following the differential factorization formula, so that the method is also applicable to the case of collinearsafe observables, i.e. to the dressed-lepton case. For this purpose, we have implemented the combination of the exact NLO QCD prediction for vector-boson production with the simulation of final-state photon radiation using PHOTOS [91]. Since we are interested in comparing to the O(α s α) corrections in our setup, we only generate a single photon emission using PHOTOS and use the same scheme for α as described in Sect. 3.1. Details on the specific settings within the PHOTOS parton shower are given in Appendix D.
In Figs. 7 and 8 we compare our best prediction (3.8) for the factorizable initial-final O(α s α) corrections to the combination of NLO QCD corrections with the approximate FSR obtained from PHOTOS for the case of W + production and Z production, respectively. For the bare-muon case also the result of the structure-function approach according to Eq. (3.13) is shown. The combination of the NLO QCD corrections and approximate FSR leads to a clear improvement compared to the naive product approximations investigated in Section 3.2. This is particularly initial-state QCD and final-state EW corrections to our best prediction δ prod×dec αsα for the case of the lepton-invariant-mass distribution (left) and a transverse-lepton-momentum distribution (right) for Z production at the LHC, as in Fig. 5. In the bare-muon case, the result (3.13) of the structure-function approach is also shown. apparent in the neutral-current process where the M distribution is correctly modelled by both FSR approximations, whereas the naive products shown in Figs. 5 and 6 completely failed to describe this distribution. In the M T,ν spectrum of the charged-current process in Fig. 7 one also finds good agreement of the different results below the Jacobian peak and an improvement over the naive product approximations in Fig. 4. The description of the p T, distributions is also improved compared to the naive product approximations, but some differences remain in the charged-current process.
In spite of the good agreement of the two versions of incorporating final-state-radiation effects, the intrinsic uncertainty of the leading-logarithmic approximations should be kept in mind. For the structure-function approach, this uncertainty is illustrated by the band width resulting from the variation (3.17) of the QED scale Q. We remark that the multi-photon corrections obtained by employing the un-expanded structure-functions Γ LL (z, Q 2 ) in Eq. (3.13) lie well within the aforementioned scale bands, which shows that a proper matching to the full NLO EW calculation is needed to remove the dominant uncertainty of the LL approximation and to predict the higherorder effects reliably. For PHOTOS the intrinsic uncertainty is not shown and not easy to quantify. The good quality of the PHOTOS approximation results from the fact that the finite terms in the photon emission probability are specifically adapted to W/Z-boson decays. The level of agreement with our "full prediction", thus, cannot be taken over to other processes.
Impact on the W-boson mass extraction
In order to estimate the effect of the O(α s α) corrections on the extraction of the W-boson mass at the LHC we have performed a χ 2 fit of the M T,ν distribution. We treat the M T,ν spectra calculated in various theoretical approximations for a reference mass M OS W = 80.385 GeV as "pseudo-data" that we fit with "templates" calculated using the LO predictions quantifying the impact of a higher-order correction in the theoretical cross section σ th is then obtained from the minimum of the function where the sum over i runs over the transverse-mass bins. Here σ th i and σ 0 i are the integrated cross sections in the i-th bin, uniformly rescaled so that the sum over all 27 bins is identical for all considered cross sections. We assume a statistical error of the pseudo-data and take ∆σ 2 i ∝ σ th i . We have also performed a two-parameter fit where the normalization of the templates is fitted simultaneously, leading to identical results. Similarly, allowing the W-boson width in the templates to float and fitting M W and Γ W simultaneously does not significantly affect the estimate of the effect of the O(α s α) corrections on the M W measurement.
In the experimental measurements of the transverse-mass distribution, the Jacobian peak is washed out due to the finite energy and momentum resolution of the detectors. In our simple estimate of the impact of higher-order corrections on the extracted value of the W-boson mass, we do not attempt to model such effects. We expect the detector effects to affect the different theory predictions in a similar way and to cancel to a large extent in our estimated mass shift, which is obtained from a difference of mass values extracted from pseudo-data calculated using different theory predictions. This assumption is supported by the fact that our estimate of the effect of the NLO EW corrections is similar to the one obtained in Ref. [50] using a Gaussian smearing of the four-momenta to simulate detector effects. The fit results for several NLO approximations and our best NNLO prediction (3.8) are given in Table 1. To validate our procedure we estimate the mass shift due to the NLO EW corrections by using the prediction σ NLOew = σ 0 +∆σ NLOew as the pseudo-data σ th in (3.18). The χ 2 distribution is shown on the left-hand side of Fig. 9 [50]. 4 Alternatively, the effect of the EW corrections can be estimated by comparing the value of M W obtained from a fit to the naive product of EW and QCD corrections (3.9) to the result of a fit to the NLO QCD cross section. The results are consistent with the shift estimated from the NLO EW corrections alone.
We have also estimated the effect of multi-photon radiation on the M W measurement in the bare-muon case using the structure-function approach given in Eq. (3.13). As discussed in detail in Ref. [46] we match the exponentiated LL-FSR corrections evaluated in the α(0)scheme to the NLO calculation in the α Gµ -scheme, avoiding double-counting. We obtain a mass shift ∆M FSR W ≈ 9 MeV relative to the result of the fit to the NLO EW prediction, which is in qualitative agreement with the result of Ref. [50].
To estimate the impact of the initial-final O(α s α) corrections we consider the mass shift relative to the full NLO result, In Ref. [50] the values ∆MW = 110 MeV (20 MeV) are obtained for the bare-muon (dressed-lepton) case. These values are obtained using the O(α)-truncation of a LL shower and for lepton-identification criteria appropriate for the Tevatron taken from Ref. [85], so they cannot be compared directly to our results. In particular, in the dressed-lepton case, a looser recombination criterion R ± γ < 0.2 is applied, which is consistent with a smaller impact of the EW corrections. Note that the role of pseudo-data and templates is reversed in Ref. [50] so that the mass shift has the opposite sign.
where M fit,NNLO prod×dec s⊗ew W is the result of using our best prediction (3.8) to generate the pseudodata, while the sum of the NLO QCD and EW corrections is used for ∆M fit,NLO s⊕ew W . The resulting ∆χ 2 distributions for the mass shift are shown in the right-hand plot in Fig. 9. In the bare-muon case, we obtain a mass shift due to O(α s α) corrections of ∆M NNLO W ≈ −14 MeV while for the dressed-lepton case we get ∆M NNLO W ≈ −4 MeV. Identical shifts result from replacing the NNLO prediction by the naive product (3.9), which is expected from the good agreement for the M T,ν -spectrum in Fig. 4. Using instead the leadinglogarithmic approximation of the final-state photon radiation obtained using PHOTOS to compute the O(α s α) corrections, we obtain a mass shift of ∆M NNLO W = −11 MeV (−4 MeV) for the bare-muon (dressed-lepton) case. The effect of the O(α s α) corrections on the mass measurement is therefore of a similar or larger magnitude than the effect of multi-photon radiation. We emphasize that the result ∆M NNLO W ≈ −14 MeV is a simple estimate of the impact of the full O(α s α) corrections on the M W measurement. The order of magnitude shows that these corrections must be taken into account properly in order to reach the 10 MeV accuracy goal of the LHC experiments. It is beyond the scope of this paper to validate the accuracy of the previous and current theoretical modelling used by the experimental collaborations in the M W measurements, which includes the O(α s α) corrections in some approximation.
Conclusions
The Drell-Yan-like W-and Z-boson production processes are among the most precise probes of the Standard Model and do not only serve as key benchmark or "standard candle" processes, but further allow for precision measurements of the W-boson mass and the effective weak mixing angle. This task of precision physics requires a further increase in the accuracy of the theoretical predictions, where the mixed QCD-electroweak corrections of O(α s α) currently represent the largest unknown component of radiative corrections in terms of fixed-order predictions.
In our previous paper [82] we have established a framework for evaluating the O(α s α) corrections to Drell-Yan processes in the resonance region using the pole approximation and presented the calculation of the so-called non-factorizable corrections. They turned out to be phenomenologically negligible, so that the O(α s α) corrections almost entirely result from factorizable corrections that can be separately attributed to production and decay of the W/Z boson (up to spin correlations).
In this paper we have presented the calculation of the so-called factorizable corrections of "initial-final" and "final-final" types. The latter were calculated in Sect. 2.4 and only comprise finite counterterm contributions which were found to be numerically very small (< 0.1%) and therefore can be safely neglected for all phenomenological purposes. The former, on the other hand, combine large QCD corrections to the production with large EW corrections to the decay subprocesses and are expected to be the dominant contribution of the O(α s α) corrections. Their calculation has been presented in Sect. 2.2, and we have shown numerical results in Sect. 3.2 for the most important observables for the W-boson mass measurement: the transverse-mass and lepton-transverse-momentum distributions for W production. The results for the neutralcurrent process comprise the invariant-mass and the lepton-transverse-momentum distributions. In the framework of the pole approximation, the only missing O(α s α) corrections are now those of "initial-initial" type. Based on the results of the NLO electroweak calculation, these are expected to be numerically small.
We have used our results for the dominant O(α s α) corrections to test the validity of simpler approximate combinations of EW and QCD corrections: Firstly, we use a naive product ansatz multiplying the NLO QCD and EW correction factors, and secondly, we approximate the O(α s α) contribution by combining leading-logarithmic approximations of QED final-state radiation with the NLO QCD corrections.
We have demonstrated in Sect. 3.2 that naive products fail to capture the factorizable initialfinal corrections in distributions such as in the transverse momentum of the lepton, which are sensitive to QCD initial-state radiation and therefore require a correct treatment of the doublereal-emission part of the NNLO corrections. Naive products also fail to capture observables that are strongly affected by a redistribution of events due to final-state real-emission corrections, such as the invariant-mass distribution of the neutral-current process. On the other hand, if an observable is less affected by such a redistribution of events or is only affected by it in the vicinity of the resonance, such as the transverse-mass distribution of the charged-current process, the naive products are able to reproduce the factorizable initial-final corrections to a large extent.
In Sect. 3.3 we have investigated to which extent the factorizable initial-final corrections calculated in this paper can be approximated by a combination of the NLO QCD corrections and a collinear approximation of real-photon emission through a QED structure function approach or a QED parton shower such as PHOTOS. For the invariant-mass distribution in Z-boson production we observe a significant improvement in the agreement compared to the naive product ansatz, since both PHOTOS and the QED structure functions model the redistribution of events due to final-state radiation, which is responsible for the bulk of the corrections in this observable. Our results can furthermore be used to validate Monte Carlo event generators where O(α s α) corrections are approximated by a combination of NLO matrix elements and parton showers.
Finally, in Section 3.4 we have illustrated the phenomenological impact of the O(α s α) corrections by estimating the mass shift induced by the factorizable initial-final corrections as ≈ −14 MeV for the case of bare muons and ≈ −4 MeV for dressed leptons. These corrections therefore have to be properly taken into account in the W-boson mass measurements at the LHC, which aim at a precision of about 10 MeV. It will be interesting to investigate the impact of the O(α s α) corrections on the measurement of the effective weak mixing angle as well in the future.
A Renormalization constants for the leptonic vector-boson decay at O(α s α)
In this appendix we provide the expressions for the finite O(α s α) counterterms to the leptonic vertices of the W and Z bosons in the on-shell renormalization scheme following the conventions of Ref. [100].
where q is the momentum carried by the vector bosons V a,b . The O(α s α) corrections to the vector-boson self-energies are given in Ref. [99] in terms of scalar functions Π V,A T . 5 Treating all quarks apart from the top quark as massless, the transverse parts of the vector-boson self-energies can be expressed as follows, The double-virtual corrections (2.16) and the (virtual QCD)×(real photonic) corrections (2.17) are obtained by dressing the virtual part of the NLO QCD corrections with the factorizable finalstate EW corrections, where the integrated counterpart of the QED dipoles I ew is defined in Eq. (2.19). Here Φ 2,IJ denotes the set of momenta of the two-particle phase space after applying the momentum mapping associated to the dipole g σ Rs⊗Rew gq b ,prod×dec = 3+γ dσ Rs⊗Rew gq b ,prod×dec − dσ Rew qaq b ,dec Φ 2+γ,(gqa)q b ⊗dV g,qa The contribution to the gq a channel is given in an analogous manner, but is not spelled out explicitly.
The collinear counterterms with additional virtual EW (2.22) and real-photonic (2.23) corrections are constructed from the corresponding term of the NLO QCD corrections by dressing them with the respective factorizable final-state corrections, 1¯ 2 dσ 0 qaq b ,PA Φ 2, 1¯ 2 (xp a , p b ) + ( 1 ↔¯ 2 ) ⊗(K + P)q a,qa The corresponding formulae for the gluon-quark channel read 1¯ 2 dσ 0 qaq b ,PA Φ 2, 1¯ 2 (xp g , p b ) + ( 1 ↔¯ 2 ) ⊗(K + P) g,qa (B.5b) and analogous expressions for the gq a channel. Here we have made the dependence on the momenta of the incoming partons explicit in order to indicate which particle undergoes a collinear splitting with the momentum fraction given by the convolution variable x.
C Non-collinear-safe observables
In order to treat non-collinear-safe observables with respect to the final-state leptons i = 1 ,¯ 2 following Ref. [89], the n-particle kinematics in the phase space of the subtraction function is treated as an (n + 1)-particle event with a collinear lepton-photon pair, where the momentum shared between the two collinear particles is controlled by the variable z iJ , where we have made explicit the cut function for the computation of observables in the notation. This modification induces additional convolution terms over the distribution [Ī ew (z)] + with which we indicate by the label "R ew ". The contribution with virtual QCD corrections is given by and analogous terms for the gq a channel. Note that the K and P operators, in general, contain plus distributions with respect to the variable x, and the above equations need to be properly evaluated in combination with the plus distribution [Ī ew (z)] + that acts on the integration variable z.
D PHOTOS settings
The results using the PHOTOS parton shower shown in Figs These settings restrict the parton shower to at most one additional photon emission in order to simulate the impact of O(α) corrections. Further settings which differ for the charged-current and neutral-current processes are as follows: W ± production: Z production: | 15,073.8 | 2014-05-27T00:00:00.000 | [
"Physics"
] |
Continuous Remote Monitoring in Hazardous Sites Using Sensor Technologies
The deployment of a distributed point source monitoring system based on wireless sensor networks in an industrial site where dangerous substances are produced, used, and stored is described. Seven essential features, fundamental prerequisites for our estimating emissions method, were identified. The system, consisting of a wireless sensor network (WSN) using photoionisation detectors (PIDs), continuously monitors the volatile organic compound (VOC) concentration at a petrochemical plant on an unprecedented time/space scale. Internet connectivity is provided via TCP/IP over GPRS gateways in real time at a one-minute sampling rate, thus providing plant management and, if necessary, environmental authorities with an unprecedented tool for immediate warning in case critical events happen. The platform is organised into subnetworks, each including a gateway unit wirelessly connected to the WSN nodes. Environmental and process data are forwarded to a remote server and made available to authorized users through a rich user interface that provides data rendering in various formats, in addition to worldwide access to data. Furthermore, this system consists of an easily deployable stand-alone infrastructure with a high degree of scalability and reconfigurability, as well as minimal intrusiveness or obtrusiveness.
Introduction
Volatile organic compounds (VOCs) are widely used in industries as solvents or chemical intermediates. Unfortunately, they include components which, if present in the atmosphere, may represent a risk factor for human health. VOCs are also found as contaminants or as byproducts of many processes, that is, in combustion gas stacks and groundwater clean-up systems. Benzene, for example, is highly toxic beyond a time-weighted average (TWA) limit of 0.5 ppm (parts per million), as compared, for instance, with the TWA limit for gasoline which is in the range of 300 ppm. Detection of VOCs at sub-ppm levels is, thus, of paramount importance for human safety and, consequently, critical for industrial hygiene in hazardous environments.
The most commonly used portable field instruments for VOC detection are hand-held photo-ionisation detectors (PIDs), which can be fitted with prefilter tubes for detecting specific gases. The pluses are that PIDs are accurate to sub-ppm levels and measurements are fast, in the range of one or two minutes; for these thus hand-held PIDs are well-suited to field use. However, they have traditionally had two drawbacks: they require skilled personnel and they cannot provide continuous monitoring. Wireless hand-held PIDs have recently become available on the market, thus overcoming these two limitations, but they have a limited battery life, in addition to being relatively costly. This paper describes the implementation and field results of an endto-end distributed monitoring system using just such VOC detectors, resulting in real-time analyses of gas concentrations in potentially hazardous sites on an unprecedented time/space scale [1].
Wireless sensor networks (WSNs), equipped with various gas sensors, have been actively used for air quality monitoring since the early 2000s [2][3][4]. WSNs have the advantage of offering full coverage of the monitored terrain by collecting measurements from redundant portions of the zone. WSNs are thus the ideal instrument for specific and efficient environmental VOC monitoring [5,6].
This paper describes the implementation of a distributed network for precise VOC monitoring installed in a potentially hazardous environment in Italy. The system consists of a WSN infrastructure with nodes equipped with both microclimatic sensors as well as VOC detectors and fitted with TCP/IP over GPRS gateways which forward the sensor data via Internet to a remote server. A user interface then provides access to the data as well as offering various formats of data rendering. The continuous monitoring, using a unique blended wired/wireless configuration, of benzene emissions from a benzene storage tank is provided as a specific example of the network's usage. A prototype of this system was installed in the eni Polimeri Europa (PEM) chemical plant in Mantova, Italy, where it has been in continuous and unattended operation since April 2011. This pilot site is testing and assessing both the communications and the VOC detection technologies.
To avoid excavations, a stand-alone system, that is, one relying only on autonomous energy and connectivity resources, was designed and installed. In terms of energy requirements, the VOC detectors proved to be by far the greatest energy user, compared to the computational and communication units. So, to ensure a sustainable battery life for these units, efficient power management strategies were studied and implemented; moreover, the WSN elements were equipped with a secondary energy source, consisting of a photovoltaic panel.
This system represents several firsts. One important novelty is the stand-alone unattended long-term operation of communications in a potentially severely hostile environment. The proprietary communication protocols being used will not be discussed in this paper whose purpose is, instead, a general description of the system. Another breakthrough concerns the sensors; the PIDs' continuous power-on operation is made possible by calibration curve linearization and by their sub-ppm detection capability.
Emission Estimation Methods
Since estimating diffuse emissions is more difficult and complex than estimating piped emissions (e.g., by stack measurement), we first established what minimum features an ideal method, for licensing and enforcement purposes, must have. The fundamental criteria we pinpointed are: (i) being inexpensive; (ii) being suitable for leak detection (all compounds, all locations); (iii) being suitable for all of the site's equipment and their phases of operation; (iv) allowing real time estimation; (v) allowing easy inspection for enforcement; (vi) allowing depiction of the emissions over any time period; (vii) allowing for reconfigurability.
A variety of methods for estimating diffuse emissions have been developed. These range from calculation to measurement, point measuring to remote sensing. Some are suited for leak detection, others for estimating annual emissions, and yet others for both of these functions. Below the main currently available methods are described, none of which, however, meets all of the criteria we identified as necessary for the ideal method.
Distributed Point Sources.
The equipment for this method consists of standard air quality measuring devices.
In order to cover all potential emission sources it is common practice to monitor several points; furthermore, instead of just fixed measuring points, mobile continuous sensors may be deployed. With the help of a "reverse" atmospheric dispersion model the emissions can be calculated from downwind air quality data combined with meteorological data. This method allows for estimating total emissions, however, it does not cover high plume emissions. Furthermore, the precise location of a leakage is hard to identify with this system.
Fixed Beam (Open Path) Optical Absorption Method.
This method measures the absorption of an electromagnetic beam (IR and UV) by gases present in ambient air, based on the principle that specific gases will absorb light from known parts of the wavelength spectra. Depending on the amount absorbed between the beam source and the detector (coupled to a spectrometer and computer) the amounts of VOCs are calculated. High plume emissions, however, cannot be measured. Moreover, the exact location of a leakage is hard to find with this method as well.
Differential Absorption LIDAR (DIAL).
Optical measuring techniques were further developed in the late nineties to overcome their limitations in pinpointing leakage and in detecting high-altitude leakage sources, resulting in DIAL (differential absorption LIDAR; LIDAR being light detection and ranging). In this system both the infrared laser beam source and the detector are located at the same end of the beam; the detector then picks up the signal from the small amount of light scattered from aerosol droplets or particles in the atmosphere. The main advantages of DIAL over fixed beam methods are that gas concentration is measured at all points along the path and no height limitations exist. Furthermore, it produces 2D and 3D maps of gas concentrations, making it possible to localise the emissions even within large industrial complexes. In other words, DIAL can estimate the total emission flux as well as localising (unexpected) leakage sources; in addition, it covers all potential emission sources (equipment, storage, loading/unloading, waste water system, etc.). However, both its accuracy of localisation as well as its differentiation between different chemical compounds are International Journal of Distributed Sensor Networks 3 limited. Nevertheless, DIAL is an outstanding complement to standard point-by-point leak detection.
Tracer
Gas. The tracer gas method consists of releasing a tracer gas (usually SF6) at different release areas and at various heights above the surface in the factory area and of measuring the VOC and tracer gas concentrations downwind of the factory via either portable syringe-based samplers or portable gas chromatographs. The emission rates of specific hydrocarbons can be estimated from simple flux reading, assuming nearby stationary wind conditions and with no significant atmospheric reactions or deposition of hydrocarbons or other release gases between the leakage points and the sampling points. We weighed each of the above methods against the criteria our system required and opted for the distributed point system for the PEM Mantova site as it combines reasonable installation and maintenance costs, reconfigurability.
Our System Overview
To craft the system, first off suitable locations (both in terms of representativeness and expected impact) were identified along the perimeter of the industrial area, along with several internal sites where hazardous emissions might potentially occur. Owing to the extension and complexity of the Mantova plant, covering some 300 acres and featuring complex metallic infrastructures, it was decided to subdivide the area involved in the piloting into 7 different subareas. Each subarea is covered by a subnetwork consisting of a sink node unit (SNU) equipped with meteorological sensors, for recording wind speed/direction and relative air humidity/temperature (eni 1 to eni 7 in Figure 1). In addition, the eni 2 unit is further equipped with a rain gauge and a solar radiation sensor. An overview of the system deployed at Mantova is shown in Figure 1.
Each SNU is connected to one or more end node units (ENUs) equipped with VOC detectors (see Figure 2 for an example of a configuration), appropriately distributed across the plant's property. This modular approach allows the system to be expanded and/or reconfigured according to the specific monitoring requirements, while providing redundancy in case of failure of one or more SNUs.
The SNUs forward meteorological data, as well as VOC concentration data, to a remote server; as noted above, Internet connectivity is provided via TCP/IP over GPRS using a GSM mobile network. Wireless connectivity uses a UHF-ISM unlicensed band. Electrical power is provided by both primary sources (batteries) and secondary sources (photovoltaic cells), as mentioned above.
VOC concentration and weather/climatic data are updated every minute. This intensive sampling interval allows the evolution of gas concentrations to be accurately assessed. Furthermore, when all of the weather-climatic measurements are collated, they provide a map of the area's relative air humidity/temperature and wind speed/direction, which are crucial for providing accurate VOC-sensor readout compensation [7]. The need for so many wind stations across the plant property is warranted by the turbulent wind distribution in that particular area, as can be observed by the different orientations of the blue arrows representing wind direction in Figure 1.
Three of the ENUs-eni 1, eni 2, and eni 3-were deployed along the perimeter of the plant to locally monitor VOC concentration while correlating it with wind speed and direction; the other seven were placed around the chemical plant and in close proximity of the pipeline, which are possible sources of VOC emissions. In fact, potential sources of VOC emissions in the plant are in easily identified areas, such as the chemical plant and the benzene tanks. The chemical plant was surrounded with a high number (6) of VOC sensors, thus creating a virtual fence, capable of effectively evaluating VOC emission sources based on the concentration patterns detected around the plant. Figure 2 shows the layout of two of the subnetworks, one deployed around the chemical plant and one near the pipeline. The subnetwork around the chemical plant, Figure 2(a), consists of two SNUs, eni 6 and eni 7, equipped with weather sensors (air/wind), each connected with three ENUs spaced tens of meters from each other. The subnetwork located in the pipeline area is shown in Figure 2(b). One of the two ENUs is located in close proximity to the end of the pipeline itself (nodo 4), while the other (nodo 2) is a bit further away. Sampling the VOC concentration at intervals of tens of meters allows the dispersion of VOC emissions to be evaluated; in addition, information about wind speed/direction allows the emission's source to be identified.
The VOC Detector
The VOC detector is a key element for the monitoring system's functionality. For this application two criteria were considered mandatory. The first is that the VOC detector must be operated in diffusion mode, thereby avoiding pumps or microfluid devices which would increase the energy requirements and make maintainability issues more critical. The second criterion was that the system should be able to operate in the very low part per billion (ppb) range, with a minimum detectable level (MDL) of some 2.5 ppb with a ±5% accuracy in the 2.5 to 1000 ppb range, which represents the range of expected VOC concentrations. Another requirement was operating at one-minute intervals.
The PID fulfills most of the above requirements. Two major issues were identified, however, which potentially impact efficient use of the PID in our system. The first was that in the low ppb range the calibration curve of the PID shows a marked nonlinearity; this would require an individualized meticulous multipoint calibration involving higher costs and complexity. The second issue was that, when operated in free diffusion mode at low ppb and after a certain time in power-off, the detector required a stabilisation time of dozens of minutes, hence it would not be able to operate at our required one-minute intervals.
Since both of the above-mentioned limitations are intrinsically related to the PID's physical behaviour, this was carefully investigated and a behavioural model of the PID was developed to explain these phenomena. To resolve the nonlinearity, a mathematical expression of the PID calibration curve was derived [8]. Accordingly, the PID calibration procedure was adjusted to measure only two parameters: the zero gas voltage and the detector sensitivity in mV/ppm. The second problem was the stabilisation time required to achieve a stable read-out time in diffusion mode at low concentrations. Originally our system requirements called for a minimum VOC data sampling interval of at least fifteen minutes. This would have had the additional advantage of prolonging the PID's battery life and maintenance, keeping costs down, However, discontinuously operating the PID results in stabilisation times of dozens of minutes in order to get a reliable PID read-out, which conflicts with the system requirement of sampling VOC concentrations in real time. In order to read very low VOC concentration levels in diffusion mode, however, the PIDs have to be continuously poweredon, consuming about 35 mA. Comparing the advantages of the two operation modes, the benefits of discontinuous operation were marginal compared to the matchless advantage of the more time-intensive monitoring of VOC concentration provided by continuous power-on operation. Accordingly, it was decided to operate the VOC detectors continuously at one-minute data sampling, which would also meet our established criteria of real-time estimation. Furthermore, this decision proved to be very effective as some emission events at the plant show very rapid variation, which would be difficult to interpret based on ten-minute sampling rates.
Continuous Emission Monitoring at Benzene Storage Tanks
Storage tanks represent an important potential source of VOC emissions and probably account for a significant amount of the site's total diffuse emissions. Thus they need to be appropriately monitored. Emissions from tanks can vary significantly from tank to tank, according to their size, design, maintenance, liquid level, and properties, as well as whether the tank is filling, stable, or emptying. Wind speed can also have a substantial effect on tank emissions, particularly for floating roof tanks.
International Journal of Distributed Sensor Networks 5 Benzene storage tanks on this site are of the floating roof type and are located in highly hazardous areas. In fact, the electrical equipment operating in those areas need a special safety certification which classifies each area according to European Directive 94/9/EC (referred to as ATEX, an acronym for the French "Atmosphères Explosibles"), regarding equipment and protective systems intended for use in potentially explosive atmospheres.
The PEM site's benzene storage tank is certified as ATEX Zone 0, which means it is an area where an explosive atmosphere is continuously present or present for long periods of time; hence, the sensors that are located in close proximity to benzene sources must meet ATEX Zone 0 requirements.
The layout of the benzene storage tank monitoring network (STMN) is displayed in Figure 3. The STMN consists of three VOC units, each equipped with a PID and a computational unit, serially interconnected by wires as well as connected to the wireless unit (WU), which provides power and wireless connectivity. The WU is then connected to the GPRS unit (eni 3 in Figure 1) which provides the Internet connectivity. The reason for choosing such a hybrid wired/wireless configuration is due to the VOC detector's energy needs.
As can be observed in Figure 3, the three VOC detectors are located in Zone 0, which requires a very high level of protection, while the WU, along with the power unit consisting of the battery and the photovoltaic panel, needed to meet the VOC detectors' high energy consumption, is located in the non-Zone 0 area. In fact, the VOC unit's current absorption of 35 mA calls for a primary energy source with at least 80 A h capacity for 60 days of continuous operation. This would mean replacing 3 batteries on the top of the storage tank, requiring skilled personnel, every two months. This was considered impractical and too costly. On the other hand, the option of equipping the unit with a secondary energy source, such as a photovoltaic panel, to prolong battery life, was dismissed as incompatible with the safety requirements of an ATEX Zone 0.
As a result, the hybrid configuration of Figure 3 was drawn up, placing the VOC units within Zone 0, yet the communication/power supply units outside of the hazardous area. This permits usage of a secondary energy source, while, at the same time, allowing easy replacement of the primary energy source.
The Communications Platform
The communications platform, described more in depth elsewhere [1], is able to support a scattered system of units collecting VOC emission data in real time, while offering a high degree of flexibility and scalability, so as to allow for adding other monitoring stations as needed. Furthermore, it provides reconfigurability, in terms of data acquisition strategies, while being more economically advantageous than traditional fixed monitoring stations.
A GSM mobile network solution featuring a proprietary TCP/IP protocol with DHCP provides Internet connectivity.
Dynamic reconnectivity strategies provide efficient and reliable communication with the GSM base station. All the main communication parameters, such as IP address, IP port (server's and client's), APN, PIN code, and logic ID, can be remotely controlled. Wireless connectivity between SNUs and ENUs is performed in an unlicensed ISM UHF band (868 MHz).
The Sink Node and Wireless Interface Units.
A block diagram of the SNU is represented in Figure 4(a). It consists of a GPRS antenna and a GPRS/EDGE quadriband modem, a sensor board, an I/O interface unit, and an ARM-9 microcontroller operating at 96 MHz. The system is based on an embedded architecture with a high degree of integration among the different subsystems. The unit is equipped with various interfaces, including LAN/Ethernet (IEEE 802.1) with TCP/IP protocols, USB ports, and RS485/RS422 standard interfaces. The sensor board has 8 analogue inputs and 2 digital inputs. The SNU is also equipped with a wireless interface (WI), shown in Figure 3(b), which provides connectivity with the ENUs.
The wireless interface (WI), Figure 4(b), provides shortrange connectivity. It operates on a low-power, ISM UHF unlicensed band (868 MHz) with FSK modulation; moreover, it features proprietary hardware and communication protocols. Distinctive features of the unit are the integrated antenna, which is enclosed in the box for improved ruggedness, as well as a PA + LNA for a boosted link budget. The PA delivers 17 dB m to the antenna, while the receiver's noise figure was reduced to 3.5 dB, compared with the intrinsic 15 dB NF of the integrated transceiver. As a matter of fact, a connectivity range in line-of-sight in excess of 500 meters was obtained, with a reliable communication with a low BER, even in hostile EM environments.
The energy required for the unit's operation is provided by an 80 A h primary source and by a photovoltaic panel equipped with an intelligent voltage regulator. Owing to its prudent low-power design, the unit can be powered with a small (20 W) photovoltaic panel while maintaining continuous unattended operation.
Given the distance among the sink nodes and the hostile EM environment, implementing a multihop WSN would have been impractical, costly and troublesome in terms of quality of service; our solution, instead, has resulted in practically flawless performance, in terms of reliability, with only a very marginal extra cost. Figure 5; it consists of a WI, similar to that previously described, yet includes a VOC sensor board as well as a VOC detector. The acquisition/communication subsystem of the ENU is based on an ARM Cortex-M3 32-bit microcontroller, operating at 72 MHz, which provides the necessary computational capability on the limited power budget available.
The EN Unit. A block diagram of the ENU is shown in
To reduce the power requirement of the overall ENU subsystem, two different power supplies have been implemented: one for the microcontroller and one for the peripheral units. The microcontroller is able to connect/disconnect the peripheral units, thus preserving the local energy resources. The VOC detector subsystem in particular is powered by a dedicated switching voltage regulator; this provides a very stable and spike-free energy source, as required for proper operation of the VOC detector itself. Communication between the ENU and the VOC detector board is based on an RS485 serial interface, providing high-level immunity to interference as well as bidirectional communication capability, which is needed for remote configuration/reconfiguration of the unit.
Network Structure and Routing Schemes.
Among the different alternatives, a hierarchical-based routing scheme was selected based on the particular nature of the installation: the extended area of the plant, the few critical areas of potential sources of emissions requiring dense deployment, and the highly uneven distribution of nodes over the area. A hierarchical-based routing scheme fit the projected deployment layout well. As said before, the installation was partitioned into subnetworks to be deployed around the critical sites, with one SNU for each individual subnetwork. In principle, wireless connectivity between the SNUs could have been implemented, using one specific SNU as a gateway to the Internet. This option, however, conflicted with at least two of the major requirements. The first is the need for redundancy in case of failure of the gateway unit; in this scenario, in fact, Internet connectivity would be lost, with consequent loss of the real-time updating capability, which is considered a mandatory requirement for the system. The second need which would not have been met is that of providing full connectivity among the individual SNUs under conditions where line-of-sight propagation was not guaranteed, due to the presence of such temporary obstacles as trucks or maintenance infrastructures. A multiple-GPRS gateway approach overcomes those limitations; even in the case of failure of one or more gateway units, Internet connectivity would be provided by the others still in operation, while the issue of the obstacles is circumvented. As for the wireless connectivity, a star configuration was preferred to a mesh configuration, given the limited number of nodes and the need to keep latency at a minimum.
Protocols and WSN Services.
Two levels of communication protocols, in a mesh network topology, were implemented. The upper level handles communications between the SNs and the server; it uses a custom binary protocol on top of a TCP layer. This level was designed and calibrated for real-time bidirectional data exchange, where periodic signaling messages are sent from both sides. Since our sensor network necessitates a stable link, quick reconnection procedures for whenever broken links should occur, were especially important. To ensure minimal data loss, the SNs have nonvolatile data storage, as well as automatic data packet retransmission (with timestamps) after temporary downlink events. Furthermore, this design is well suited for low-power embedded platforms like ours, where limited memory and power resources are available. In fact, our protocol stack currently requires about 24 KB of flash memory (firmware) and 8 KB RAM. The lower level, in contrast to the upper one, concerns the local data exchange between the network nodes. Here a cluster tree topology was employed; each node, which both transmits and receives data packets, is able to forward packets from the surrounding nodes when needed. In this specific application, the topology and routing schemes are based on an ID assigned to each EN unit, where the ID can be easily adjusted using selectors on the hardware board. This choice allows for easy support and maintenance, even when nonspecialized operators have to install, reinstall, or serve one or more units.
Energy Budget Issues
Energy budget plays a key role in the maintainability of the WSN [9]. In our case this is made even more critical by the necessity for stand-alone operation, as well as due to periodic maintenance intervals in excess of four months. Since electrical energy from the plant could not be used, secondary sources had to be locally available; photovoltaic panels (PVP) fit the bill. The SNUs are almost all equipped with PVPs, as they have to support a number of functions, including connectivity and data collection from sensors. The ENUs, when equipped with low-energy sensors, have 3 to 5 years of battery life using primary sources [10]. However, in this system the ENUs have to support the power-hungry VOC sensors as well. For this reason, the ENUs are also equipped with PVPs.
ENU Energy Budget.
The EN nodes have been fully deployed since July 2011; since that time we have noticed that the VOC sensors' energy budget predominates over that of the computational/communication unit. Since the PIDs used for reading the VOC concentration need to be continuously on to operate efficiently, this corresponds to a current draw of some 35 mA, corresponding to 840 mA h a day, more than twice the amount used by the communication/computational units, which consume some 360 mW a day. The ENU's primary source capacity is 60 A h, which provides more than 2 full months of continuous operation.
To enable the ENUs to rely fully on autonomous energy resources while providing continuous operation, a 5 W photovoltaic panel was integrated into the unit in order to supply the additional 360 mW average required power. The PVP, however, can fulfil the task only under ideal sunlight conditions, that is, in summer, but hardly at all in winter. Hence, the PVP power supply unit also has a charge regulator which was specifically designed to provide maximum energy transfer efficiency from the panel to the battery regardless of weather conditions. In conclusion, the secondary energy source plays a key role in ensuring the stand-alone and unattended operation of the communication platform.
As a concrete example, Figure 6 shows the battery voltage plots for the ENUs connected to SNUs 5 and 6. As can be observed, the ENUs exhibit quite satisfactory charge conditions, although ENU 4 (eni 5 nodo 4) shows a slightly lower voltage level, probably due to deployment in a partially shadowed area. The battery voltage remains above a 12.3 V value, with a slightly decreasing trend, possibly due to lower solar energy given the fact that the time period corresponds to the onset of autumn.
SNU Energy
Budget. The SNUs were deployed at the PEM plant in the middle of April 2011. They have much higher energy requirements than the ENUs as they have to supply energy for both wireless connectivity and sensor operation (including meteorological sensor).
The average current draw is around 90 mA, corresponding to a power consumption of about 1 W. SNUs have superior primary and secondary resource capabilities, with a 2-month battery life relying on the primary source alone. Figure 7 shows the battery voltage for the eni 2 to eni 7 SNUs; the eni 1 plot is missing as the chart can only represents 6 graphs in each diagram. As can be observed, all the units demonstrate a minimum voltage exceeding 12.3 V, which denotes a satisfactory charging state. In this period there is also a slight decrease in the minimum battery voltage value, showing an energy imbalance between the primary and secondary sources, mostly due to sunlight reduction, as in the graph for the ENUs.
Detailed information about charge status and trending are also available; Figure 8 gives an example of how the current drawn by or supplied to the battery is compared with charge status. As shown here, the energy balance keeps the battery voltage at a steady, satisfactory level. Extensive data logs and reports are available to help the maintenance team in evaluating any critical event or servicing required to keep the system in full operation.
VOC Detector Energy Budget.
Continuous power-on operation requires a 35 mA h charge, which corresponds to about 2 months of full operation with a 60 A h primary energy source. However, as described above, this can be lengthened due to PVP integrated power. Furthermore, since the UV lamp's expected life is more than 6000 hours of continuous operation, that is, about four months, routine maintenance for the system-UV lamp replacement, PID refitting and battery replacement-can be planned on a fourmonth cycle.
Experimental Results
Data from the individual sensors deployed on the field can be directly accessed and presented in various formats. First, the data from the field is forwarded to a central database for storage; then, for effective data rendering the system offers a web-based interface, allowing us to process and view data in real time. There are several formats for displaying the main factors, such as weather information and VOC concentrations. However, it is also possible to access raw data, and generate summary reports relating to specific time periods and specific network areas. Furthermore, all monitored variables may be related to each other visually.
Following are examples of the various formats the interface offers.
The overview map (Figure 9(a)) provides information about the entire network installed at the PEM plant. Each SNU is represented with a gray circle and a blue arrow indicating the wind direction. By selecting one of the circles (Figure 9(b)) a summary panel appears that lists the current temperature, humidity, wind speed and direction, as well as VOC concentrations over a daily, weekly, or monthly period. Minimum and maximum values of the day are also shown.
International Journal of Distributed Sensor Networks These representations, obtained using an interpolation algorithm, are in pseudo-colour. Blue denotes a lower concentration/temperature, while red indicates a higher one. This format is particularly useful for quickly identifying quickly if there are any areas within the plant with high VOC concentrations and whether they are related to meteorological conditions. It should be emphasized that the choice of red was merely a chromatic one; it has absolutely no reference to any risky or critical condition.
An additional visual rendering of the data gathered by each sensor on the field can be obtained by opening the graphic panel window, see Figure 11. This panel allows anyone to display graphically the stored data in any time interval; up to six different and arbitrarily selected sensors can be represented in the same window, for analysis and comparison. The graph in Figure 11 shows the trend in VOC concentration values detected by the six PIDs deployed around the chemical plant over one month; the background values are similar to each other, demonstrating the effectiveness of the calibration procedure.
By selecting specific areas within the chart it is possible to further increase the details of the graph, allowing an easier interpretation of any short peak detected. In fact, thanks to the intensive 1-minute sample interval, the evolution of the concentration, along with other relevant weather parameters, can be accurately displayed. Moreover, a useful inspection tool allows users to quickly record min, max and mean values in any selected period. Figure 12(a) shows VOC concentration readings over a long term (4 months) from a sensor positioned along the perimeter of the industrial area. As can be observed, the data reveals an increase in the background value during the summer, probably due to higher temperatures; in fact, the peak value (a concentration greater than 500 ppb), registered around July 25th, was due to meteorological conditions that affected the dispersion of pollutants, as seen in Figure 12(b). This graph combines the peak VOC value with a very low wind speed, conditions which favor the accumulation of pollutants from the nearby production plant. However, this was merely a local increase since the other sensors simultaneously recorded much smaller peaks. Furthermore, the values decreased from September on. Figure 13 shows the trend of VOC concentrations in the pipeline area. Two sensors are positioned in the same area, but VOC 1 is closer to the emission sources and consequently has a higher background reading. Both sensors clearly demonstrate the cyclical effects of day and night.
When VOC sources need to be identified, the correlation between wind/speed direction and VOC concentration is vital; for this reason, a graphic representation relating these two parameters was mandatory. In Figure 14 different representations of VOC concentrations combined with the wind direction data are shown for four detectors located along the plant's perimeter; this particular plot shows the wind directions referenced to the north and the VOC concentration in ppb. ENU 1, located in the southwest corner of the plant, is shown in Figure 14. VOC concentration is higher in quadrants I and IV, showing that the net VOC flow is entering the plant area. This may be related to the emissions generated by the traffic on the motorway running along the west side of the plant, or it may be due to emissions from other industrial sites, such as the petroleum refinery (WSW) and its storage area (WNW) located just across the motorway. In the same diagram, ENU 2, which is instead located at the northwest corner of the plant, the highest concentration values are in quadrants I and II; also in this case the VOC flow is coming from north of the plant area. In contrast, the other plots (near ENU 5) show a homogeneous distribution of VOC concentration from all directions. Hence, this format of data rendering allows PEM plant staff to quickly identify and visualize wind directions in areas of high levels of VOC concentration, while giving an overview of the predominant orientation of the VOC flux during the day.
The system's interface also allows users to keep tabs on specific areas, as in the case of the sensor network installed in the proximity of the benzene storage tank roof. In fact, data gathering was useful in identifying those phases in processing that are potentially more significant in terms of emissions. Specifically, concentration peaks are encountered with regular frequency (see Figure 15) after completely filling the tank; since, in such conditions, the floating roof is located closer to the sensors, eventual sealing problems along the perimeter of the roof due to wear or warping of the seals can be discovered. Furthermore, in the summertime the high volatility of the compounds could lead to variations in VOC concentrations during the emptying of the tank as well. Whatever the situation, the registered values, however, always are within the accepted range, since they are from a direct emissions source.
Conclusions
An end-to-end distributed monitoring system of integrated VOC detectors, capable of performing real-time analysis of gas concentration in hazardous sites at an unprecedented time/space scale, has been implemented and successfully tested at an industrial site near Mantova, Italy. The fundamental criteria for our system were providing the site with: a flexible and cost-effective monitoring tool that identifies emission sources in real-time year round, using easily redeployable and rationally-distributed monitoring stations that were suitable for all of the site's equipment, in order to achieve better management of abnormal situations.
Piloting the system allowed us to pinpoint additional key traits. For example, collecting data at 1-minute time intervals meets several needs: identifying short-term significant events, quantifying the emission readings as a function of weather conditions as well as of operational procedures, in addition to identifying potentially hazardous VOC sources in the plant area.
Moreover, the choice of a WSN communication platform gave excellent results, above all in allowing for redeploying and rescaling the network's configuration according to specific needs as they arise, while, at the same time, greatly reducing installation costs. Furthermore, real-time data through a web-based interface provided both adequate levels of control and quick data interpretation in order to manage specific situations. The program offers multiple formats for visualizing the data. In terms of the actual detectors, among the various alternatives available on the market, PID technology proved to meet all the major requirements, as PIDs are effective in terms of energy consumption, measuring range, cost, and maintenance once installed in the field. The energy budget was another significant element to be considered, particularly in ATEX Zone 0 areas, so secondary sources (PVPs) were adopted. Finally, fitting weather sensors at the nodes of the main network stations ensured a clearer understanding of on-field phenomena and their evolution, thus providing accurate identification of potential emission sources.
Future activity will involve, primarily, standardizing the application to allow deploying the WSN in other network industries (e.g., refineries), in addition to assessing WSN infrastructure monitoring of other environmental indicators. | 8,783.6 | 2012-07-01T00:00:00.000 | [
"Engineering",
"Environmental Science"
] |
A Domain Extension Algorithm for Digital Error Correction of Pipeline ADCs
A domain extension algorithm to correct the comparator offsets of pipeline analog-to-digital converters (ADCs) is presented, in which the 1.5-bit/stage ADC quantify domain is extended from a three-domain to a five-domain. This algorithm is designed for high speed and low comparator accuracy application. The comparator offset correction ability is improved. This new approach also promises significant improvements to the spurious-free dynamic range (SFDR), the total harmonic distortion (THD), the signal-to-noise ratio (SNR) and the minor analog and digital circuit modifications. Behavioral simulation results are presented to demonstrate the effectiveness of the algorithm, in which all absolute values of comparator offsets are set to |3Vref/8|. SFDR, THD and SNR are improved, from 34.62-dB, 34.63-dB and 30.33-dB to 60.23-dB, 61.14-dB and 59.35-dB, respectively, for a 10-bit pipeline ADC.
Introduction
ADCs are widely used in many areas, such as music recording, healthcare, radar systems and communication [1].A trend of the modern ADC design is the use of digital background calibration to compensate for the raw performance of analog circuits [2][3][4][5][6][7][8][9].However, many digital background calibrations can only correct gain errors [10,11], which are caused by finite op-amp gain and capacitor mismatches.This leaves the comparator offsets corrected by the traditional digital error correction technique or not corrected at all.The traditional 1.5bit/stage ADC can only correct the comparator offsets within ± ref 4 V [12].For small-geometry transistors, typical mismatches in the width, length and threshold voltage can lead to significant comparator offsets [13].
Comparator offsets greatly limit the accuracy of a switched capacitor pipeline ADC.In this paper, a new algorithm is developed to improve the comparator offset correction ability for the 1.5-bit/stage pipeline ADC.This innovative algorithm increases the comparator offset toleration ability by 50%.In addition, the algorithm also provides crucial information on both overflow and underflow situations.
Domain Extension Algorithm
The residue plot of a real ADC with comparator offsets is shown using dashed lines.In this case, the maximum comparator offset is ref 4 V and the corresponding maximum output equals to ref V .Since the output of the current stage is the input of the next stage, and the input range is from ref V , the out of range output leads to code loss.In order to prevent the ADC from code loss, the comparator offsets should be within the range of with an added overflow/underflow judgment.Two Matlab behavioral simulations are used to illustrate the improvement of the comparator offset correction ability for the proposed ADC.The first ADC behavioral simulation includes eight traditional 1.5-bit/stage converters followed by a flash ADC, and the second ADC behavioral simulation includes eight trial 1.5-bit/stage converters also followed by a flash ADC.In these simulations, the absolute values of the comparator offsets are set between 0 and ref 0.5V .In order to control and narrow research findings, all 1.5-bit/stage ADCs are onlycomplicated by the comparator offsets.In addition, the flash ADCs setting are ideal.In these simulations, the total number of conversions is 14 2 .The total miscode numbers, and their related comparator offsets, are show in , for the ADC based on the proposed algorithm.
The transfer function of the traditional 1.5-bit/stage pipeline ADC is given by the following equation [14]: . ( The transfer function of the proposed five-domain 1.5bit/stage pipeline ADC is given by the following equation: This proposed ADC consists of eight 1.5-bit stages followed by a 2-bit flash ADC.There are 12 total output bits, 10usable bits and the first two bits are utilized as overflow/underflow bits.Figure 3 shows the algorithm process.In order to have the digital output of five-domain 1.5-bit/stage ADC consistent with the traditional 1.5-bit/stage ADC, the subtraction of one operation is needed.Since 000 minus 1 is negative, adding a "1" in front of the digital output of the first stage avoids the negative number.For the same reason, the later stages also need to subtract one operation.In addition, the dislocation addition should be implemented before the subtraction of one.The first two bits are overflow/underflow bits.Therefore, when they are "11" or "01", they will reference to the input signal beyond or below the reference The circuit level implementation of Equation ( 2) is given by , 000 , 001 , 010 , 011 , 100 In Equation ( 3), the two capacitors are equal.When the required gain is one, the circuit level realization is the same as the traditional technique, and capacitor 2 C connects to the corresponding reference voltage.However, a gain of two for ref V cannot be realized through the traditional technique since one of the capacitors is the feedback capacitor.The maximum gain for ref V is the non-feedback capacitor divided by the feedback capacitor, which is one.To extend the domain, a new method is proposed.In this new method V dd need to be set to twice the ref V .The first and the last equations of (3) are Φ is high, the converters work on the sample phase, input is sampled on the two capacitors simultaneously.When 2 Φ is high, they work on the amplification phase, the feedback capacitor 1 C connects to the output and the non-feedback capacitor 2 C connects to the corresponding reference voltage.
The proposed algorithm slightly modifies the analog.Two comparators are added to extend the quantify domains, and two references are used to provide a gain of two for ref V .Since the actual configuration is fully dif-
Simulation Results
In order to demonstrate the effectiveness of the domain extension algorithm, a 10-bit pipeline ADC was simulated in MATLAB.The ADC consisted of eight fivedomain 1.5-bit/stage converters and a 2-bit flash ADC.In the simulation, all absolute values of comparator offsets were set to
Conclusion
The decrease of the transistor geometry causes problematic mismatches in width, length and threshold voltage, which leads to significant comparator offsets.These comparator offsets, in turn, greatly limit the performance of ADCs.However, the traditional digital error correction technique can only correct the absolute value of comparator offsets lower than ref 4 V .Therefore, in order to improve the comparator offset toleration ability, a domain extension algorithm has been presented, which can correct the absolute value of comparator offsets within ref 3 8 . This new approach involves minor analog and digital modifications and increases the comparator offset toleration ability by 50% with overflow/underflow judgment.Simulation results have revealed significant improvements of SFDR, THD and SNR performance.
Figure 1 (V − and ref 4 V
Figure 1(a) shows the residue plot of the traditional 1.5bit/stage ADC.In Figure 1(a), two ideal threshold voltages are ref 4 V − and ref 4 V shown with dotted lines.The coded range is from ref
Figure 1 (Figure 1 .
Figure 1.(a) Residue plot of the traditional 1.5-bit/stage ADC and (b) residue plot of the five-domain 1.5-bit/stage ADC.range is from
Figure 2 .
According to Figure 2, for the ADC based on the traditional digital error correction technique, miscodes occur when the absolute values of the comparator offsets are higher than ref 4 V .By comparison, no miscodes occur for the absolute values of the comparator offsets lower than ref 3 8 V
Figure 4 (.
c) shows the processing of output codes based on the traditional digital error correction technique with the comparator offsets the same as Figure 4(a), but the threshold voltages are Figure 4(d) shows the processing of output based on the traditional technique with the comparator offsets set to be zero.These three cases, Figures 4(a), 4(b), and 4(d), have the correct digital output, while Figure 4(c) is different from the other three because the traditional technique cannot correct the absolute values of comparator offsets higher than |V ref /4|.
)Figures 5
Figures 5(a) and 5(b) are the circuit configurations based on the traditional technique and the proposed algorithm, respectively.ref V is simplified by r V in the figures.Although the actual configurations are fully differential, the sing-ended the configurations are shown for simplicity.When 1Φ is high, the converters work on the sample phase, input is sampled on the two capacitors simultaneously.When 2 Φ is high, they work on the amplification phase, the feedback capacitor 1 C connects to the output and the non-feedback capacitor 2 C connects to the corresponding reference voltage.The proposed algorithm slightly modifies the analog.Two comparators are added to extend the quantify domains, and two references are used to provide a gain of two for ref V .Since the actual configuration is fully dif-
Figure 4 .Figure 5 .
Figure 4. (a) Example of the proposed algorithm with the comparator offset of 3V ref /8; (b) Example of the proposed algorithm with no comparator offset; (c) Example of the traditional algorithm with the comparator offset of 3V ref /8; (d) Example of the traditional algorithm with no comparator offset.
was set to 45-MHz, and sample rate was set to 100-MS/s.The Fast Fourier Transform (FFT) plot of this simulation using the traditional method is shown in Figure6(a).The dynamic performance as shown in the FFT plot is recorded as 34.62-dBSFDR, 34.63-dB THD, and 30.33-dBSNR.
Figure 6 ( 7 . 8 VFigure 6 .
Figure 6.FFT plots of a 10-bit pipeline ADC using (a) the traditional digital error correction technique and (b) the proposed algorithm.
Figure 7 .
Figure 7. Simulated dynamic performance using (a) the traditional digital error correction technique and (b) the proposed algorithm. | 2,149 | 2014-02-25T00:00:00.000 | [
"Computer Science",
"Engineering"
] |
Beyond the Warring States: The First World War and the Redemptive Critique of Modernity in the Work of Du Yaquan (1873–1933)
The intellectual impact of the First World War in China is often understood as having led to a disenchantment with the West and a discrediting of the authority of “science”, while at the same time ushering in a renewed sense of cultural as well as national “awakening”. Important developments such as the May Fourth Movement, the rise of Chinese Marxism, and the emergence of modern Confucianism have become integral parts of the narrative surrounding the effects of the “European War” in China, and bear witness to the contested relation between tradition and modernity in twentieth-century Chinese thought. Through a case study of a number of wartime and post-war texts written by the “cultural conservative” thinker and publicist Du Yaquan (1873–1933), this paper tries to draw attention to the complexity and occasional ambiguity of responses to the “Great War” in modern Chinese intellectual history. More specifically, the following pages offer an analysis of Du’s critique of “materialism” in the context of his quest for social freedom and cultural continuity, his enduring commitment to scientific notions of social evolution and political governance, and his approach to the relations among war, the nation-state, the individual, and the international interstate order developed against the background of the First World War.
Introduction: The "Great War" in China as Event and Narrative
There is an oft-quoted saying by the French poet and essayist Paul Valéry (1871Valéry ( -1945 according to which the First World War confronted humanity with the fact that civilizations too are mortal beings (Valéry 1977, 94). 1 In the context of the intellectual history of modern China, it might be more accurate to say that in the wake of the war, Chinese thinkers learned that Western civilization in particular was mortal, if not already moribund. This at least is how the story was and still is often framed: the post-war period in China was one of national as well as cultural "awakening" (juewu 觉悟) (see Wang 2016, 41-48), and entailed a call for nothing less than a "liberation from the West" (Zheng 2011). 2 Generally speaking, the discourse surrounding the impact of the First World War on China hinges on fluid terms such as "civilization" and "culture", and draws heavily on dramatic metaphors of "death", "awakening", and "rebirth". Perhaps this already indicates that the war does not figure so much as a factual event in this context, but rather as a narrative structure, one allowing for a decoupling as well as recombination of discursive elements from historically and culturally distinct traditions, at least on a more abstract level. 1 What is usually ignored however is that Valéry's melancholy diagnosis is followed by a celebration of the "European genius" in the second part of his text.
2 More precisely, Zheng Shiqu 郑师渠 understands such "liberation" as coinciding with an end of the normative appeal of capitalism and the rise of historical materialism, as if the social reality of the war had opened up the cracks in the ideological superstructure of the New Culture Movement necessary for Chinese Marxism to impose itself.
Admittedly, the horror of trench warfare, massive civilian casualties, and unimaginable destruction during the "Great War" may seem to rail against the adoption of such a dispassionate approach. However, we are not, in my view, merely dealing with a stubborn indifference to the cruelty and contingency of historical events which always threaten to shatter the crystal palace of philosophical abstraction. In retrospect, we can clearly see that the brutal reality of armed conflict did not prevent Western as well as Chinese thinkers from approaching the struggle between the great powers as an opportunity for reassessing their respective traditions as well as the prospects for a possible encounter or reconciliation between them. In turn, such a rethinking was seen as a response to very real and pressing socio-political issues. After all, as the historian James Q. Whitman claims, in the modern conception of war, armed conflicts are supposed to deliver a "verdict", in the sense that "victory in war either proves or legitimates a certain cultural, moral, or metaphysical value" (De Warren 2014, 727).
To be sure, the many problems besetting the embattled nations were widely reported in Chinese media (Sachsenmaier 2007, 118), even if the First World War seems not to have been primarily approached from a "phenomenological" standpoint focused on the lived experience of soldiers and civilians on the frontlines by most Chinese thinkers. Travel journals and the reports of Chinese living in Europe at the time and published after the war contain detailed eyewitness accounts which offer a more personal and lively counterweight to the somewhat dreary and repetitive discourse on the "Decline of the West" often associated with this period. 3 As Eugene W. Chiu 丘为君 indicates, while the Chinese experience of the "European War" (Ouzhan 欧战), as it still sometimes referred to in China, was at first characterized by a certain detachment, the mass of reports and analyses in journals and newspapers allowing the events on the Western front to be approached as a gargantuan "text", Chinese commentators gradually shifted their attention to the actual living conditions of common people caught up in the war (Chiu 2005, 94, 118).
Just as importantly, many if not all intellectuals in China were highly concerned with how the situation in Europe would impact the East-Asian context, especially after Japan (aided by Great Britain) started moving in on Germany's concessions in Shandong province. As such, they were hardly unaware of the global dimension and broader geopolitical implications of what was, after all, an increasingly worldwide conflict. What is crucial to point out, however, is that more philosophically minded observers approached the war not so much as a factual occurrence, but rather from a more macroscopic perspective, that is to say, as an epochal event (in a quasi-Badiouian sense) necessitating an "awakening" and a retrospective insight into its larger historical and cultural causes and conditions. China's definitive loss of Shandong to Japanese imperialist ambitions following the Versailles Peace Conference of 1919 obviously played an important role in this respect.
As Du Yaquan's 杜亚泉 (1873-1933 statement which serves as an epigraph to my paper indicates, the causes and conditions of the First World War were not necessarily sought in the recent past alone. For Du, chief editor of the influential journal Eastern Miscellany (Dongfang zazhi 东方杂志) between 1911 and 1920, 4 the social and ideological upheaval characteristic of the modern era could in some sense be seen as entailing a return to the political chaos and intellectual confusion (or, in positive terms, richness and ferment) of the Warring States period (481-221 BCE) in Chinese history. 5 As anyone familiar with the development of traditional Chinese philosophy knows, such an identification should not only be read in a negative sense, since this period is also the origin of the "hundred schools" of pre-Qin thought. More to the point, as Nicolas De Warren notes with respect to the philosophical response to the war in Europe, it is easy to forget that when the First World War broke out, it was also greeted with a certain sense of enthusiasm by some thinkers, as an event harbouring the potential for a social revolution and "destructive renewal" of the world within itself 4 Du had de facto already been in charge of the journal's affairs since 1909, see Wang (2016, 5).
5
The analogy between the Warring States period and the modern world order following the end of the Qing dynasty and the collapse of the tianxia 天下 ("all-under-heaven") paradigm became an even more prevalent theme during the Second World War with the appearance of the so-called "Warring States Faction" (Zhanguo ce pai 战国策派), a group of intellectuals (most of them Tsinghua University graduates) associated with the bimonthly journal Zhanguo ce 战国策, which was published in the beginning of the 1940s and was followed up by an eponymous supplement to the Chongqing-based newspaper Dagongbao 大公报. Common themes in the writings of "Warring States" intellectuals were a reappraisal of the philosophy of Nietzsche (and German culture in general), a tone of militarist nationalism, and a defence of "hero worship". He Lin 贺麟 (1902Lin 贺麟 ( -1992, often credited with having been the first to use the expression "New Confucianism", was also counted among the ranks of the "Warring States Faction". For more information, see Fung (2010, 120-26). A representative figure of this relatively short-lived current of thought, which came to be condemned as "fascist" on the mainland after the founding the People's Republic, was the Shakespeare specialist Lin Tongji 林同济 (1906-1980, in whose article "The Recurrence of the Age of the Warring States (Zhanguo shidai de chongyan 战国时代的重演)" many of the themes mentioned in the above are joined together. In this text, Lin makes it clear that the idea of the "Warring States" refers to a universal phase in the history and socio-cultural evolution of different societies (each culture having a distinct Gestalt, tixiang 体相). As such, it denotes a stage of total warfare (quantizhan 全体战), where every single thing and person is mobilized for the sake of war, a process Lin sees as being epitomized by the Qin dynasty which unified China at the end of the Warring States period in 221 BCE. For Lin, war was thus not something to be solved or prevented, but rather embraced as a means for the self-assertion of the Chinese nation (see Lin 1983, 443-44).
(see De Warren 2014, 716). 6 Likewise, in China, figures as diverse as the radical intellectual Chen Duxiu 陈独秀 (1879-1942 and the more moderate and reconciliatory Du Yaquan saw the Great War as a tragic manifestation of the patriotism of the citizens of European nations. As such, it was also an opportunity to reflect on what they perceived to be the lack of patriotic spirit among their compatriots and raise the Chinese nation from its state of slumber and stagnation (see Zheng 2011, 70-71;Zhang 2016, 113). 7 As Du wrote, in biologistic terms which I will further explore below, the mind of organisms is always stimulated and aroused to action by impressions coming from its surroundings. The same applies to the people of a country (guomin 国民). Our self-absorbed and protective compatriots have remained in a state of stagnation for thousands of years due to a lack of stimuli from the outside world. (Du 1914b, 187) Additionally, there was a perhaps surprising amount of Germanophile sentiment among Chinese intellectuals after the war broke out, at least until China officially declared war on Germany in 1917. Contributors to the flagship journal of the New Culture Movement New Youth (Xin qingnian 新青年), such as Chen Duxiu saw the Germans as a "springtime people" (qingchun zhi guomin 青春之国民), whose cultural energy they contrasted with that of older and "decaying" European nations, most notably France, as the birthplace of a revolution that had failed to make good on its promises and normative demands on a global scale (see Zhao 2017, 109-12;Zhang 2016, 112).
In more general terms, a relatively positive appraisal of the intellectual impact of the war is still seen among contemporary Chinese observers. The Taiwanese scholar Edward W. Chiu, for instance, presents the Great War as a veritable catalyst for an "Enlightenment" in China (Chiu 2005). The mainland Chinese historian Zheng Shiqu 郑师渠 has argued that these dramatic historical events allowed the West to overcome an arrogant and exaggerated belief in the merits of its own civilization, while at the same time freeing Chinese thinkers from decades of self-depreciation and feelings of cultural inferiority (Zheng 1997, 213-14). Similarly, Xu Guoqi, a historian who has done much to draw attention to the neglected role of China in the First World War, characterizes the latter as a "vehicle for China's 6 Some scholars believe that the First World War played a considerable role in the already emerging rift between continental and analytical philosophy, and served as a catalyst for the closely related decline of British Idealism after the latter's German Idealist sources fell into disrepute. (See Vrahimis 2015, 84-93, andMorrow 1982) 7 A few months after the armistice, Du wrote a short article outlining the various "benefits" (liyi 利 益) China had gained during the conflict in predominantly pragmatist terms (Du 1919b).
transformation, renewal, and regeneration" (Xu 2005, 10). As he puts it, "the war provided the momentum and the opportunity for China to redefine its relations with the world through its efforts to inject itself into the war and thus position itself within the family of nations" (ibid., 9). While such arguments are probably intended to be descriptive rather than ideological, it should at the same time remind us of the importance of carefully considering in what sort of narrative the Chinese response to the war is framed and retold. According to Dominic Sachsenmaier, already at the time "a variety of groups in China, from free-trade liberalists to early Marxists (…) saw the Great War as part of a teleological history" (Sachsenmaier 2007, 120). In Xu Guoqi's opinion, the ultimate explanation behind China's apparent eagerness to join the war effort is to be found in what he calls the Chinese "obsession" (Xu 2005, 2) with joining the ranks of the international order, an attitude which supposedly also conditioned the overall response of Chinese intellectuals to the outbreak of the war.
However, if we direct our attention to analyses of the cultural-historical trajectory seen as leading up to the war, specifically those made by thinkers critical of (Western) modernity, a less clear-cut picture imposes itself. 8 More precisely, Xu Guoqi's assessment seems to underestimate the extent to which reflections on the war were not only about an imagined and long-awaited convergence between China and the West, and were not merely focused on the prospect of China finally coming into its own as one nation-state among others, but also gave rise to more ambiguous and at times incongruous reflections on the nature and limits of modernity and its political institutions. The intention of this paper is to highlight and explore some of these ambiguities in the writings of Du Yaquan, who is usually labelled as a cultural conservative without further examination of to what degree this is actually true. Before turning to a more detailed analysis of Du's philosophical reflections on the "Great War" in relation to the question of Chinese modernity, I will proceed by first providing some additional background information that will allow us to get a better picture of the broader cultural impact of the First World War on Chinese intellectual history.
Post-war Chinese Discourse on Science and the Shifting Boundaries of the "New"
The above observations indicate that the Chinese response to the Great War, in which China participated as a "forgotten ally" (Alexeeva 2015) 9 supporting the Allied Forces by dispatching an estimated 140,000 Chinese labourers to the Western Front, 10 has to be framed in a larger historical context. The two Opium Wars and China's defeat at the hands of Japan in the First Sino-Japanese War in 1895 had already made it clear that the waning Qing empire needed to adopt modern (especially military) technology. With the increasing implausibility of maintaining a rigid conceptual distinction between a Chinese "substance" (ti 体) and a Western "function" or "application" (yong 用), the adoption of technology was gradually discovered not be a mere matter of "technique" (shu 术) as opposed and inferior to "learning" (xue 学), but to involve the appropriation of "science" (gezhixue 格致学, later kexue 科学 (see Elman 2004)) as well. In this context, "science" was understood not so much as a mathematized form of objective inquiry, but rather as a much more generally applicable and socially performative "method" and "spirit" (see Luo 2000, 57-66) that would allow China to successfully achieve modernization and position itself in the world as a sovereign nation. As Wang Hui 汪晖 has aptly put it, science thus took on the form of a veritable "moral imperative" (Wang 1989, 23).
Moreover, modernization was seen as something that not only had to occur on an institutional and political level, but also on that of individual virtue, not in the least by radically reinterpreting the relation between the "private" sphere of morality and the "public" domain of politics, a view epitomized by Liang Qichao's 梁启 超 (1873-1929) call for the creation of a "new citizen" or "new people" (xinmin 新 9 For Olga Alexeeva (2015, 44), the fact that the design for a grandiose mural entitled Panthéon de la guerre, commissioned by the French State while the war was still ongoing as a celebration of all allied nations and their contributions to the envisaged victory, originally included Chinese labourers, only to be replaced by the figures of American soldiers in the final version, symbolizes the fact that the Chinese war efforts were consigned to oblivion in Western historical consciousness.
10 See Xu 2005, 114-54. The Republic of China adopted a strategy known as "labourers in the place of soldiers" (yigong daibing 以工代兵), labourers which were recruited and dispatched to Europe through the intermediary of private companies, thus allowing China to retain a semblance of neutrality while still supporting the Allied Forces against Germany. This strategy was devised by Liang Shiyi 梁士诒 (1869-1933), a cabinet minister and a close confidant of Yuan Shikai. Liang, sometimes dubbed the "Chinese Machiavelli", had already started arguing for the strategic importance of China entering the war at the side of the Allied Forces in 1914. He saw it as a way for China to achieve full recognition as a nation-state, not in the least through a return of German concessions in Shandong. (See ibid. 82-83, 87, 90-91) Ironically, most of the Chinese labourers sent to the frontlines were recruited from Shandong province, which was later ceded to Japan at the Paris Peace Conference.
民). 11 The growing awareness of the need for science, as the blueprint for culture as a whole, is usually understood as coinciding with an increasing loss of the normative power of the Chinese tradition, particularly of Confucianism, as a model for political governance, communal life, and individual conduct. The failure of the newly founded and politically unstable Chinese Republic to prevent General Yuan Shikai from proclaiming himself emperor in 1915, a move that was backed by Kang Youwei's 康有为 (1858-1927) "Confucian Religious Society" (Kongjiao hui 孔教会) which proposed installing Confucianism as a state religion, further fuelled calls for the abolishment of traditions seen as inhibiting the emergence of a "new culture" (xin wenhua 新文化) and to what the intellectual historian Luo Zhitian 罗志田 has termed a "worship of the new" (Luo 2017, 1-60).
Within this familiar synoptic account, the period following the First World War is usually interpreted as signalling a shift away from this "worship of the new" and a naïve celebration of all things Western toward a more conflicted and at times syncretistic approach to what became known as the "problem of Eastern and Western cultures" (dongxi wenhua wenti 东西文化问题). 12 As far as Du Yaquan for instance was concerned, the war had endowed the seemingly straightforward yet highly changeable and indeterminate terms "old" and "new" with a completely different sense. In his view, the "new", which had previously more or less meant imitating the West, now had to give way to a different kind of "novelty", that is to say, to the creation of a genuinely "new" form of culture that would not simply coincide with a one-sided emulation of Western civilization, but combine elements of the "new" and the "old" within itself (see Du 1919c, 401-2). Just as importantly, after the war "the West" ceased to be seen as a consistent totality, but instead began to appear as a force-field of contradictory if not antagonistic forces (see Luo 2017, 250-51). The spectacle of advanced technology being put to the service of relentless slaughter and destruction had caused science to be "put to shame by the cruelty of its applications" (Valéry 1919, 97). In turn, the continuity between "science" and "democracy", as symbols for the epistemological and institutional requirements of modern society (and quasi-religious objects of faith in the discourse of the New Culture Movement, see Wang 1989, 22-23) was ruptured, in the sense that scientific and technological ingenuity had clearly failed to translate into a rational organization of individual societies and the international order as a whole (see Han 2017). Instead, a gaping chasm had opened up between "force" (li 力) and "principle" (li 理) (Zhang 2016). The reputation of the sort of social Darwinism previously embraced by many Chinese thinkers suffered considerably in the process (Xu 2018, 163). Additionally, Western philosophers associated with German militarism became symbols of the malaise of modernity and prominent targets of critique. 13 In a lecture entitled "The Crisis of European Culture and the Direction of China's New Culture" (Ouzhou wenhua zhi weiji ji Zhongguo xin wenhua zhi quxiang 欧洲文化之危机及中国新文化之趋向) from 1922, Zhang Junmai 張君 勱 (1887-1969) went so far as to claim that continuing to slavishly emulate Western nations after the war would signify the end of culture (wenhua 文化) as such, since there would no longer be any "patterns/refinement" (wen 文) or "transformation" (hua 化) (Zhang 1922, 238) in the first place. 14 To be sure, although it is tempting to be carried along by the sweeping statements many intellectuals made at the time, some nuance and restraint is necessary in this context. This much Zhang Junmai actually indicates himself a little further on in the text of the same speech, when he argues against making simplistic overgeneralizations concerning Western and Chinese cultures. A similar caution should be displayed when it comes to the supposed discrediting of science in post-war China. It is often claimed that the destruction and suffering brought on by the war put a definite end to the optimistic belief in science, the most well-known example undoubtedly being Liang Qichao's call to awaken from the "dream of the omnipotence of science" following his tour of Europe between 1919 and 1920 (see Zheng 2006).
However, what Wang Hui has called the "community of scientific discourse" (kexue huayu gongtongti 科学话语共同体)-a community extending beyond the "scientific community" in the narrow sense, thus including all intellectuals who invoked concepts derived from scientific reasoning or articulated their views by appealing to the discourse of science-managed to far outlive such largely rhetorical attacks. Wang argues that the two world wars did not in fact end up undermining the authority of science, quite to the contrary: this competitive world scene reinforced sovereign states' demands for science and technology, further guaranteeing the development of science and technology, professionalization, state control of science and technology, and the dominant position of the scientific worldview. (Wang 2008, 131) In his view, this dominant position is also reflected in the influential "debate on science and metaphysics" from 1923, a debate in which "metaphysicians" such as Zhang Junmai and Liang Qichao argued for maintaining the proper boundaries between scientific and humanistic modes of reasoning and cast doubt on the applicability of a scientific outlook to the domains of "existence", "morality", "culture", and "politics", as distinct fields of knowledge and action irreducible to "science". As Wang Hui emphasizes, the position of the "metaphysical" camp was thus not that of an outright rejection of science, but rather reflected an implicit acceptance of the scientific attempt to arrive at a rational division of labour and functionally differentiated taxonomy of knowledge across fields of learning which could no longer be reconstituted into a coherent whole or an unmediated continuum (see ibid., 132-37).
Crucially, questioning the "omnipotence" of science in the context of the postwar "awakening" to its limitations and pathological consequences almost never came down to a straightforward call for the restoration of traditional forms of knowledge, but rather entailed a shift toward an assertion of the importance and autonomy of other, equally novel fields of knowledge, such as "philosophy". 15 This much becomes apparent in the following passage from an article Zhang Dongsun 张东荪 (1886-1973 Here, "science" and "philosophy" have already become universally applicable categories of knowledge that are no longer constrained by geography, culture, or time and are explicitly framed in relation to the equally universalist desideratum of social freedom (a "commonwealth" instead of a "dictatorship"). Following the abandonment of traditional Chinese taxonomies of knowledge, it would be these universalized terms that would serve as vehicles for the reassertion and renegotiation of cultural particularity. Additionally, we should bear in mind that, at least to some extent, Chinese post-war critiques of science and "Western materialism" echoed the Romantic self-critiques of many European intellectuals at the time (see Zheng 1997, 213;Sachsenmaier 2007, 111). As such, they should not be confused with indiscriminate assaults on Western culture as a whole, but can rather be seen as creative appropriations and reconceptualizations of such auto-critiques. 17 The post-war European interest in Chinese "wisdom", or the "wisdom of the East" in general, undoubtedly influenced the attitude of Chinese intellectuals toward their own tradition as well. 18 What is also important to remember is that such reappraisals of the value of Chinese culture were not always met with a warm welcome in China. Some like the liberal pragmatist Hu Shi 胡适 (1891-1962) feared that the protests directed at Western power politics and the perfectly justified critiques of the atrocities of the Great War would degenerate into a renewed Chinese sense of "arrogance" and "complacency", the Orientalist admiration for China expressed by some Western scholars in his view merely counting as a "temporary psychopathological state" (quoted in ibid., 210).
In any case, as the title of Zhang Junmai's lecture quoted in the above indicates, what was at stake for Chinese thinkers in their reflections on the war was both 17 Henri Bergson (1859Bergson ( -1941, one of the thinkers most often invoked by the "metaphysicians" in their critique of scientism during the 1923 debate, was involved in propagandist denunciations of "the mechanization of spirit" (Bergson 1915, 36) he associated with Prussia/Germany and in drawing binary distinctions between the "élan vital" of the French people and the mechanistic materialism of Germany. Similarly, on the German side, the vitalist philosopher Rudolf Eucken , another favourite of the Chinese "metaphysicians", approached the war as a means for the liberation of Germany and German culture.
the "crisis" of Western culture as well as the development of a "new culture" for China. The adoption of a civilizational discourse in which a wedge was driven between "novelty" or "modernity" on the one hand and "the West" on the other was a means of articulating this ambiguous and unstable position. In the process, "conservative" critics of "Western" modernity tried to wrest equally "Western" ideologies such as Marxism and socialism from their cultural confines and redefine them as genuinely universal political projects that could draw on, or be reconciled with, the Chinese tradition. As Du Yaquan for one insisted, after the war the "old" Europe had to give way to a new civilization propelled by the rebirth of the "old" culture of China in combination with a "new" (i.e. non-militarist) Western culture. Hence, it is not so surprising to find the "supposedly conservative" Du Yaquan declaring the lower classes of all countries to be the true subjects and victors of the war, and greeting the rise of international socialism with much enthusiasm. In his view, it is only from the perspective of the "old world" of militarism where "right is might" that the end of the war and a farewell to its "instruments of misfortune" (不祥之凶器) 19 could count as defeat instead of a liberation (Du 1919a, 206-8).
Du believed the abolition of class differences and economic inequalities to be the only sure means to put an end to military conflict once and for all (see Du 1914b, 191;Du 1918e, 458). His position thus hardly shares anything in common with a straightforwardly conservative withdrawal into already discredited political and ethical models without any regard for the structural features and ideological discourse of modern societies.
The post-war "problem of Eastern and Western cultures" gave rise to heated debates between radical iconoclasts and more moderate thinkers who still believed in the viability of certain aspects of the Chinese tradition. However, both shared a mistrust of the Western powers following the "betrayal" of the Versailles Peace Treaty, which led to student demonstrations and strikes across the whole of China, ushering in what later became known as the May Fourth Movement. As such, they shared a common concern over "culture" (wenhua 文化, Kultur), and not merely "civilization" (wenming 文明, Civilization), that is to say, a form of "awakening" and "enlightenment" that would, in one way or another, reflect and serve the particularities of China as a nation, regardless of whether these particularities were understood in a culturally determinate or a more universalist sense (see Xu 2018).
After the Versailles "betrayal", cultural conservatives had to abandon the notion that Woodrow Wilson's League of Nations counted as an incarnation of the age-old Confucian idea of datong 大同 ("great unity") (see Xu 2005, 253-54). Nor could Chen Duxiu still speak, as he had done in the period of short-lived enthusiasm immediately following the German defeat, of a "victory of universal principle over power" (公理战胜强权, or, more colloquially: "the victory of right over might") (quoted in Gao 1999, 9). Instead, Chen had come to terms with the fact that any "universal principle" always remains dependent on the support of political and military power, without which it would remain an easy prey for the powers that be (see Chen 1982). Clearly, then, following the war, both radicals as well as conservatives were engaged in a pursuit of the "new", that is to say, a different kind of "novelty", the semantic horizon of which had expanded considerably in the meantime. 20 20 In this respect, it is worthwhile considering the work of Ku Hung-Ming (Gu Hongming) 辜鸿铭 (1857-1928), born in the British colony of Penang (in Malaysia) and educated in Edinburgh, who is usually portrayed as the epitome of an arch-conservative "reactionary" and a living fossil from Imperial China. However, a closer examination of one of his books, The Spirit of the Chinese People, which bears the Chinese subtitle Chunqiu dayi 春秋大义 (The Great Meaning of the Spring and Autumn Annals) from 1915, partly written in response to the American missionary Arthur Henderson Smith's (1845-1932) (in)famous Chinese Characteristics from 1894, which had remained popular in the first decades of the 20th century, quickly complicates the picture. The Spirit of the Chinese People contains a lengthy appendix entitled "The War and the Way Out" (Ku 1915, 147-68) which is interesting to consider in the present context. The importance Ku attached to this essay is apparent from the fact that he already provides a summary of his main argument in the preface to the whole book, which has the ambition of showing his readers the "real Chinaman" and the actual "characteristics" of Chinese civilization. While Ku claims that Chinese civilization is now in a position to "save" the war-torn West, his staunchly "conservative" line of reasoning is full of praise for Germany, which he sees as "the true, rightful, and legitimate guardian of the modern civilization of Europe" (ibid., preface, 15). While he concedes that German militarism is the immediate culprit for the outbreak of the war, Ku argues that the German "worship of might" should actually be seen as a reaction against the "religion of mob-worship" (the subtitle of his essay) he associates with British civilization in particular. As he puts it later on in the main text of the essay itself: "If there is to be peace in Europe, the first thing to be done, it seems to me, is to protect the rulers, soldiers and diplomats from the plain men and women; to protect them from the mob, the panic of the crowd of plain men and women which makes them helpless." (ibid., 154) He then goes on to argue that the German (over)reaction against "mob worship" can be balanced out and remedied by returning to a Confucian "religion of good citizenship", that will allow nations to expect absolute loyalty from their subjects, thus giving rise to a "Magna Carta of loyalty" (see ibid., 9-12). Additionally, in Ku's view, the "mob-worship" on the level of politics had been exacerbated by the "mob rule" of the commodity in the "selfishness and cowardice" of what he calls "the spirit of Commercialism" (see ibid., preface, 18-19). For Ku, then, the problem that surfaced with the war was not the rupture between "science" and "democracy", or an excess of "Westernization", but rather the delirious influence exerted by the "mob-worship", as represented by democratic politics and the capitalist economy, on Western civilization as a whole. While his position clearly contain elements which are straightforwardly identifiable as "conservative", his radical reinterpretation of Confucianism as simply amounting to a "religion" that can ensure loyalty to the state confronts us with the unwieldiness and indeterminacy of the term "conservatism" in modern Chinese intellectual history which Benjamin I. Schwartz already identified decades ago.
Du Yaquan on War, Materialism, Evolution, and Statehood
In the remainder of this paper, I will attempt to provide more concrete illustrations of the general observations made in the above by analysing a number of Du Yaquan's wartime and post-war writings that are indicative of the complexity of the cultural conservative Chinese response to the First World War. In doing so, I will start by considering the socio-political dimension and significance of his critique of "materialism". Although this type of anti-materialism may at first sight appear to be a hackneyed and predictable theme echoing the cliché of a "spiritual East" versus a "materialist West", we should bear in mind that it continued to figure prominently in later Republican-era "debates" (literally "wars of opinions/discourses", lunzhan 论战), namely those on "science and metaphysics" (1923), the applicability of historical materialism and its categorization of the developmental stages of society to Chinese history (from the late 1920s to early 1930s), and the conceptual validity of dialectical materialism vis-à-vis formal logic and science (during the first half of the 1930s). Moreover, as I will try to show in what follows, post-war cultural conservative attacks on "materialism" are not to be dismissed out of hand as reactionary gestures drawing on a simplistic and culturalist East-West dichotomy, but have to be understood as part of an intellectual effort to rethink the modern normative requirement of social freedom.
Du Yaquan almost immediately started paying close attention to the "European War" and contributed a significant number of articles to this topic in Eastern Miscellany, which became one of the journals providing the most extensive and detailed coverage of the war under his editorial leadership (Chiu 2005, 95-98;Wang 2016, 54). Du wrote a series of reports (xuji 续记) on the latest state of affairs concerning the war from 1914 to 1917, which were later collected in a slim volume entitled A History of Events in the European War (Ouzhan fasheng shi 欧战 发生史) published by Shanghai Commercial Press in 1924 (see Chiu 2005, 103). However, it is not these factually oriented and largely descriptive texts, but rather his philosophical analyses of the underlying causes behind the war as well as the latter's broader cultural significance for which Du is still remembered to this day. In a particularly well-known text, entitled "The State of Our Compatriots' Awakening After the End of the Great War" (Dazhan zhongjie hou guoren zhi juewu ruhe 大战终结后国人之觉悟如何) from 1919, Du makes it clear that the war has led to an awareness of the necessity of spiritual as well as material reform on a global level (Du 1919a, 205). 21 In other words, he is not simply proposing a reassertion of the dominance of "spirit" over "matter" along the lines of Rabindranath Tagore's (1861Tagore's ( -1941 triumphalist praise for the putative spiritual superiority of Asia as a whole. In effect, one of the most interesting aspects of Du's writings is the coexistence of culturalist and universalist orientations, which are not always easy to disentangle. Thus, while Du famously described the West as a "dynamic civilization" as opposed to a "static" China, insisting that this is not merely a gradual but a substantial difference, he at the same time took care to note that the lives of a considerable portion of the Western populace were still entirely "static" in nature. Employing the universalist distinction between the urban and rural as metaphors for the tension between tradition and modernity, Du compared his compatriots' pre-war blind admiration for the West to the situation of a farmer or shepherd from the countryside who is dazzled by the hustle and bustle of city life without being aware of all the contradictions and social suffering there (see Du 1916c, 343). As Feng Youlan's 冯友兰 (1895-1990) (see Van den Stock 2016, 144-52) and Liang Shuming's 梁漱溟 (1893-1988 socio-political philosophy (see Van den Stock, forthcoming) as well as the development of Maoism bear out, the deceptively simple binaries of traditional-modern, Chinese-Western, and rural-urban would give rise to varied and by no means straightforward conceptual constellations throughout the subsequent history of modern Chinese thought.
The abovementioned tension between "culture" and "civilization" can also be found in Du Yaquan's critique of materialism. Already a year before the war broke out, Du published a series of three essays bearing the title "On Saving the Nation through Spirit" (Jingshen jiuguo lun 精神救国论) in Eastern Miscellany. However, in contrast to what the title might suggest, Du does not engage in an indiscriminate attack on the philosophical position of materialism here, but rather targets the latter more selectively and strategically, namely by engaging in an extensive critical overview and discussion of evolutionary theory and social Darwinism. These reflections are explicitly articulated against the background of the rise of European colonial militarism, which Du portrays as an incarnation of the "animal nature" unleashed by the "materialist" view of the world as a struggle for power in which might is right. Du argues that the "materialist" pursuit of "wealth and power" (fuqiang 富强) and lopsided interpretations of the theory of evolution (tianyan 天演) 22 were introduced into China at a time when their adverse social consequences had already begun to become evident in the West and a resurgence of "idealist" positions could begin to be discerned (Du 1913, 33-34). In this context, Du explicitly links "idealism" with a certain voluntarism, that is to say, a belief in the power of human autonomy and self-determination. In contrast to 22 Du is obviously referring to Yan Fu 严复 here. Incidentally, the war led to a volte-face in Yan's own attitude toward Western culture and the Chinese tradition at large. (See Luo 2017, 251) "materialism", Du saw an urgent need for the pursuit of a social freedom that departs from the irreducibility of the human being and its spiritual-moral capacities. That "idealism" is a very fluid category for Du becomes clear from the fact that it is supposed to includes thinkers as diverse as Montesquieu, Hume, and Hegel. Another indication of Du's association of idealism and materialism with autonomy and heteronomy, respectively, can be found in his analysis of the authoritarian turn in Japanese politics following the Russo-Japanese War of 1904-1905, which Du sees as reflecting a departure from an "idealist" belief in the power of the human mind that was still embraced at the beginning of the Meiji Restoration (see ibid., 37-38).
What Du Yaquan proposes over and against the immoral kind of "materialist" evolutionary theory that had cast the modern world into a merciless struggle for the survival of the fittest is what he calls "social cooperationism" (shehuixielizhuyi 社会 协力主义) (Du 1915a), a notion inspired by the anarchist Peter Kropotkin's (1842Kropotkin's ( -1921 idea that "mutual aid" plays an important role in biological as well as social evolution. After the war, Cai Yuanpei 蔡元培 (1863-1940) would also describe the victory of the allied forces as coinciding with a triumph of Kropotkin's ideas over "militarist" Nietzscheanism and social Darwinism (see Cai 1984, 203). Crucially, for Du, "cooperationism" also points toward a future synthesis between nationalism and internationalist pacifism. In his view, such a synthesis had become unavoidable given the increasing economic interdependence between nations in a world governed by military and monetary power (see Du 1918c). Even more importantly, a reconciliation of nationalism and internationalism would ideally serve to prevent events such as a world war from ever happening again. However, invoking the transition from "governing the state" (zhiguo 治国) to "pacifying all-under-heaven" (ping tianxia 平天下) prescribed in the classical Confucian text of The Great Learning (Daxue 大学), Du argues that any future form of "internationalism" would have to be grounded in a prior cooperation between citizens on the level of the nation-state (see Du 1915a, 21-22). The need for attaining a balance between "strength" (jianqiang 坚强) as well as "reconciliation" or "harmony" (tiaohe 调和), as quasi-cosmological concepts Du primarily deploys in analysing the "phenomenal" (youxing 有 形) dimension of politics, would first of all have to be realized "internally", that is to say, inside of a certain nation-state and people, before the latter can attempt to peacefully position itself within an international interstate order (see Du 1916b, 171-73). In short, in the same sense that "inner" moral perfection is the precondition for "outer" social order in the traditional Confucian logic of governance, nationalism counts as the logical precondition for internationalism here.
In Du Yaquan's view, while China had traditionally been preoccupied with "governing" (zhi 治), that is to say, ensuring the general well-being of its own people, and thus remained relatively indifferent to the possible existence of other states falling outside of the scope of "all-under-heaven" (tianxia 天下), it now had to come to terms with a more competitive world-order in which "protecting" (wei 卫) the nation had imposed itself as a new and urgent political imperative, all while remaining on guard against a form of militarism that would depart from China's supposed tradition of pacifism (see Du 1915c). 23 In his own words: "our compatriots should become aware that the existence of the state is a factual and not a conceptual affair, and that its basis of existence is located in military power, and not governance through culture (wenzhi 文治, more colloquially: 'civil administration')" (Du 1915c, 149). Interestingly enough, Du associates what he takes to be the traditional Chinese focus on "internal governance" (neizhi 内治) with an attitude of indifference toward the external material world supposedly found in "Indian contemplative philosophy (印度之潜心哲学)" (ibid., 148). This again indicates that his attack on "materialism" has little or nothing in common with "idealism" as it is defined in more vulgar examples of Marxist intellectual historiography. In contrast, Du's "idealism" is profoundly activist in orientation and serves as a means of safeguarding the possibility of autonomy in the face of historical processes which are beyond the control of individual human beings. In this sense, before all else, "spirit" serves a symbol for autonomy rather than denoting a specific metaphysical position. Indeed, for Du, the problem lies not so much in the analytical privilege given to the tangible aspects of human existence by materialist theories, but rather in the very notion of ideology and its imposition of misleading abstract requirements on social reality, as the passage below vividly illustrates: Those who are now propagating various "isms" seem to be begrudge the integratedness (tongzheng 统整) of our traditional culture and cannot refrain from engaging in manoeuvres to acquire power and luxurious wealth, using Western thought as a pretext to bring it to ruins (…) Expecting to be saved by these various isms would be like expecting the devil to show us the way into paradise. Oh you demons, the end is upon you! (魔鬼乎，魔鬼乎，汝其速灭)" (Du 1918b, 367) Despite his frequent appeal to the Chinese tradition, then, Du Yaquan's position cannot be straightforwardly identified as "conservative", and does not entail 23 By contrast, in his wartime private correspondence Yan Fu favoured a much more pragmatic and utilitarian approach, in which any moral and normative considerations would have to be temporarily subordinated to the task of saving the nation. Yan argued that China needed to return to the military strength and vigour of the Qin dynasty and the strategic acumen of the Legalist school of pre-Qin philosophy, rather than focus on moral supremacy. Additionally, his observations of the "European War" had led him to the conclusion that the democratic system was hardly conductive to the efficient mobilization of military force. (See Chen 2012, 122-23) a rejection of the new political form of the nation-state, but rather involves a complex attempt to mediate between tradition and modernity. This is precisely why the term "reconciliation" (tiaohe 调和) figures so prominently in his writings on the "problem of Eastern and Western cultures". Du's repudiation of social Darwinism is a case in point, since he continues to work under the assumption that there is a strong parallelism and even a continuum between nature and society and that the same force or constellation of forces govern the domains of the physical and the social. The use of physiological metaphors of "anaemia" and a symptomatic "excess of blood" in his post-war diagnosis of the condition of a "static" China and a "dynamic" West (see Du 1916c, 342) already suggests as much. 24 These biologistic metaphors obviously call to mind Chen Duxiu's call to reinvigorate the "metabolism" of the Chinese body politic with the cells of a new culture and remove its old and "rotten" elements in A Call to the Youth (Jinggao qingnian 警告青年) from the inaugural issue of New Youth in 1915 (see Chen 1915). 25 As Du himself put it unambiguously with reference to the question as to whether the current situation of a world embroiled in war can really be blamed on individual states or political parties: "That which governs the tendencies in the world of society is actually no different from the natural forces governing the ten thousand things." (Du 1917c, 194) This also becomes apparent in a text from 1916, where Du describes the war in cosmological terms as an embodiment of the tension between "love" (ai 爱) and "strife" (zheng 争) (Du 1916a). Evolution in both the natural and the social world is thus approached as the result of an interplay between contradictory forces such as the centripetal and centrifugal forces in physics (cf. Du 1916b;1918a). While such an approach seems to shift the burden of accomplishing a transformation of society from the individual to history as a process that escapes the immediate control of nations as well as citizens, Du's "anti-materialist" leanings leave the door open for the individual (and by the same token, the state) to regain command of its own fate.
The cosmological appropriation of the logic of evolutionary theory sketched in the above has important consequences for understanding Du Yaquan's approach to the reconciliation of nationalism and internationalism he envisaged against the backdrop of the Great War. Again, for all of his criticism of the social Darwinist sort of "evolution without ethics" (to paraphrase the title of Huxley's famous book), Du clearly embraces the basic logic of evolutionary thinking in arguing that the progression from a state of savagery to one of civilization involves a change in the reasons for which war is fought: from a broader historical perspective, Du discerns a progression from the rationale behind warfare which moves from contingent empirical reasons (i.e. immediate bodily needs, in which case wars remain on the level of struggles between animals or squabbles between children), to a calculated consideration of "interests" or "benefit and harm" (lihai 利害), to finally reach a point where normative and ideological considerations enter the fray, and wars are fought over "right and wrong" (shifei 是非), such as for example the American Civil War (see Du 1915b). 26 Within this line of reasoning, the Great War counts as an archetypal "war of ideas" (sixiang zhan 思想战) over right and wrong, and is not merely a battle between conflicting, unreflective animal instincts. In this sense, we can say that Du's "conservatism" is one which has already internalized certain "scientific" narratives of historical development that were far from discredited through the event of the war. Rather, the latter provided him with an opportunity to rethink and redeploy these narratives, all while attempting to link them with elements from the Chinese philosophical tradition.
We should bear in mind here that Du started his career as an autodidact intellectual who devoted himself to introducing natural scientific knowledge into China after having abandoned the prospect of pursuing a career as a scholar-official after reaching the entry-level degree of xiucai 秀才 ("flowering talent") in the imperial examination system at the age of 16. Eight years later, in 1898, Du was recruited by Cai Yuanpei, the future president of Peking University who then still served as rector of the Shaoxing Chinese-Western School (Shaoxing zhongxi xuetang 绍兴中西學堂) in Zhejiang, to become a teacher in mathematics, meanwhile applying himself to the study of natural scientific subjects such as chemistry, mineralogy, zoology, as well as philosophy, politics, and other "humanist" disciplines, which were more likely still seen as part of the epistemological continuum of what Neo-Confucian thinkers called the "investigation of things" (gewu 格物). In 1900, Du founded an academy for the study of science in Shanghai and published the inaugural issue of Yaquan zazhi 亚泉杂志, one of the first Chinese journals devoted to popularizing the natural sciences (with a focus on chemistry), to which he would personally contribute a significant number of texts and translations (from the Japanese) until it ceased publication in 1901. Du invested much of his time and sometimes his own resources in science research and education, as well as to such mundane affairs as a setting up a shop selling laboratory equipment in Shanghai. His endeavours as an author and editor at Shanghai Commercial Press led to the publication of pioneering works such as the Comprehensive Botanical Dictionary (Zhiwuxue da cidian 植物 学大辞典) (1918) and the Comprehensive Dictionary of Zoology (Dongwuxue da cidian 动物学大辞典) (1923) (see Xie 1988, 8-11). Tellingly, Du's activities toward the spread of scientific knowledge hardly stopped after the war (see Chen, Kang, andYao 2008, 1046-49 Crucially, Du Yaquan's appeal to the authority of "science" also surfaces in his critique of the deficiencies of modern democracy. In Du's view, the majority of the common people want as little as possible to do with politics and remain completely indifferent to the affairs of the state. The Chinese people's overall apathy and lack of knowledge makes them questionable subjects of the "awakening" necessitated by the Great War. As Du put it: "the so-called will of the people is actually so somnolent as to appear involuntary (所谓民意者，实则为 朦胧无意而已)" (Du 1917c, 195). In this sense, it seems that Du expected "Mr. Science" to come to the aid of "Mr. Democracy": in his utopian vision of a future where all nations and military factions will be abolished and national democracies will give way to global socialism, he imagined the emergence of a new social class that would combine specialized scientific knowledge with the practical skills and energetic potential of the labouring population. This activist class of scientists would serve to supplant the apathetic unconscious "will of the people" largely driven by "material" desire instead of rational choice (ibid., 198). 27 Du's cosmological-evolutionary framework for the interpretation of natural as well as social changes provides us with an important clue to the significance of what he defended as the outlook of "continuism" (jiexuzhuyi 接续主义) (Du 1914a). From a "continuist" perspective, there is no necessary contradiction between the old and new or tradition and modernity. In socio-political terms, this means that the continuation of the past into the present does not come down to a reactionary attitude aimed at restoring an already defunct social order, but rather embodies a unity of conservatism and progressivism ensuring that national unity is not only safeguarded on a spatial-territorial, but also on a temporal-historical level. As Wang Hui has shown, the questions of national sovereignty and cultural continuity were closely connected in Du's writings (Wang 2016, 60), where the "reconciliation" of the old and the new is presented as being predicated on such a "continuist" attitude. Interestingly enough, whereas Du's 1914 text on continuism written just before the outbreak of the war still called for subordinating the individual to the interests of the state, in his wartime and post-war writings, the nation-state begins to appear as the medium for the reconciliation of opposites, that is to say, as a place where the dialectical interplay between the cosmological forces of the centripetal ("love") and the centrifugal ("strife") as well as the opposition between the private and the public could be balanced out.
In an article from 1917 entitled "On the Boundaries between the Individual and the State" (Geren yu guojia zhi jie shuo 个人与国家之界说), Du came to argue that individualism should be reconciled with, and not sacrificed to, nationalism, a position he takes up in opposition to German militarism (Du 1917a, 168). Such a reconciliation involves drawing the proper boundaries between the domain of the individual and that of the state, instead of propagating a straightforward subordination of individual to national interests. At the same time, he assumed that upholding these boundaries could also serve the purpose of preventing individual interests from usurping the public good. Once again, Du's argument is framed within the Confucian logic of the continuity between individual self-cultivation and the governance of the state, with Du invoking a passage from the Analects (14.42) which insists on the necessity of "cultivating oneself in order to bring peace to the common people (修己以安百姓)" (quoted in Du 1917a, 167). His line of reasoning thus wavers between the two poles he seeks to reconcile and takes up an ambiguous position in between individualism and nationalism. Du proposes that if individuals are simply sacrificed for the sake of the nation without being given the opportunity to "cultivate themselves", they would in effect cease to be of any use to the state, since they would have no proper self or "personality" (renge 人格) to sacrifice in the first place. In his own words: "if we want people to fully devote themselves to the affairs of the state, we have to first allow them to care for themselves" (ibid.). In his view, top-down government measures have to be supplemented with a "spiritual socialism" (精神上之社会主义) (Du 1919a, 210) on the level of individual morality. It is not clear if this should be read as a defence of individualism per se, or merely as a functionalist argument in which individuals must be allowed to develop themselves for the greater good of the state. On the one hand, Du seeks to reaffirm the traditional continuity between "governing the self " (zizhi 自 治), that is to say, moral autonomy, and "governing the state" (zhiguo 治国), while at the same time insisting on the importance of upholding the proper boundaries between state and individual. In short, Du seems to be struggling here with what would continue to be a dominant theme in retrospective evaluations of the New Culture and May Fourth Movement, namely the conflict between the search for "national salvation" and the pursuit of "enlightenment" (i.e. individual autonomy) as Li Zehou 李泽 厚 famously, if rather simplistically, put it (see Li 1987). In spite of his rejection of Li's diagnosis, a very similar conclusion was reached by Gao Like 高力克, who argued that in these movements of "unfinished enlightenment", "'individual awakening' was merely an indirect manifestation of 'national awakening'" (Gao 1999, 11). Du Yaquan's historically informed writings on the relation between individual and state can thus be seen both as a precursor to more recent Chinese discourse on the "dialectics of Enlightenment", as well as a possible resource for comparative philosophical reflections on the possibility of social freedom in the modern world.
Conclusion
The examples given in the above indicate that Du Yaquan did not seek to repudiate, but rather to redeem modernity, as something containing the potential for a "reconciliation" between the past and present, as well as contradictory aspects of intrastate and interstate politics within itself. As such, Du's critique of evolutionism could go hand in hand with an analysis of war as a quasi-natural catastrophe, one destined to eventually evolve into a vehicle for the attainment of political freedom and economic equality on a global level. Similarly, his condemnation of the economic injustices he saw as the basis of the Great War was accompanied by a strong belief in the ability of industrial capitalism to continue increasing productivity, while redirecting the latter toward the creation of actual material wealth and disentangling it from unequal relations of distribution (see Du 1918e, 459). Perhaps most importantly and timely from our current perspective, in analysing the Great War Du explicitly called for critically reflecting on the limitations and dangers of nationalism, an ideology he tended to present as a necessary evil rather than a positive good, and, paradoxically, as the only means available to China to secure a position within a more long-term historical process leading to the overcoming of the nation-state (see Du 1917d, 398;1918d). Rather than simply being concerned with the relation between individual and state in general, the problem for Du would seem to have been that, under the condition of the continuing threat of war and the ever-present possibility of a return to an age of "Warring States", it is the "individuality" of the state within a competitive global order of nation-states which provides the basis for individual human well-being and the right to subsistence. Within this logic, there is no space of mediation and "reconciliation" between the individual (private) and the social (public) in the absence of the nation-state. Lacking the necessary cohesion and resistance against external aggression, China would become violently assimilated into the economic realm of Western colonialism and lose its autonomy to unbridled and goalless "material" impulses, thus effectively falling back to a more atavistic, pre-normative stage in the evolution of society and being severed from the necessary "continuist" connection to its own tradition. As such, for Du, the "individuality" of the state comes before that of the individual in the strict or ordinary sense, precisely because war has consistently threatened to undercut the already fragile social and moral cohesion of the Chinese people throughout its modern history. While not going as far in his critique of the category of the nation-state as contemporary Chinese intellectuals who advocated reasserting the traditional notion of "all-under-heaven" (tianxia), Du's conflicted attitude toward nationalism is testament to the modern dialectics of autonomy, where the requirements of freedom and autonomy are always caught in a tension between the spheres of the individual, the state, and geopolitical interstate conflicts. By contrast, invocations of the ideal of "all-under-heaven" as a straightforward alternative to the "Western" notion of the state conveniently ignore the fact that the logistics behind the realization of a universalist vision such as that of tianxia risk remaining caught up in the geopolitical logic of modernity, that is to say, one of different nation-states ruthlessly competing for the benefits of global capitalism, as the only de facto universality in the contemporary world. Over a century after the armistice, Du's wartime and postwar writings remind us of the fact that relation between intrastate political freedom and interstate war is not an extrinsic one, and that the historical specificity of this relation should not be left out of the picture in comparative political thought. | 13,867 | 2021-05-07T00:00:00.000 | [
"History",
"Philosophy",
"Political Science"
] |
Enthalpy Effect of Adding Cobalt to Liquid Sn-3.8Ag-0.7Cu Lead-Free Solder Alloy: Difference between Bulk and Nanosized Cobalt
Heat effects for the addition of Co in bulk and nanosized forms into the liquid Sn-3.8Ag-0.7Cu alloy were studied using drop calorimetry at four temperatures between 673 and 1173 K. Significant differences in the heat effects between nano and bulk Co additions were observed. The considerably more exothermic values of the measured enthalpy for nano Co additions are connected with the loss of the surface enthalpy of the nanoparticles due to the elimination of the surface of the nanoparticles upon their dissolution in the liquid alloy. This effect is shown to be independent of the calorimeter temperature (it depends only on the dropping temperature through the temperature dependence of the surface energy of the nanoparticles). Integral and partial enthalpies of mixing for Co in the liquid SAC-alloy were evaluated from the experimental data.
INTRODUCTION
Unique physical−chemical properties and microstructure features make Sn−Ag−Cu (SAC) alloys the worldwide most used lead-free solders. In particular, the Sn-based Sn-3.8Ag-0.7Cu (wt %) alloy (SAC387), which corresponds to Sn-4.1Ag-1.3Cu (at. %), is employed extensively in the modern electronics industry. 1−3 However, two main problems caused by using such type of lead-free solder are still not solved (i.e., the much higher melting temperature compared to traditional lead-containing solders and the extensive growth of brittle intermetallic layers). During the past decade, many attempts were made to decrease the melting temperature of lead-free SAC solders and improve the mechanical reliability of the corresponding solder joints. 4,5 One of the most popular ways to achieve such improvements is the addition of a fourth element. 6−9 According to investigations of the mechanical and thermodynamic properties and the microstructure, minor doping of active nanoparticles should be a promising solution of the above-mentioned problems. 10−12 Different heat effects are expected for the addition of Co in bulk and nanosized form into the liquid SAC387 alloy, and this difference is caused by the surface enthalpy of the nanoparticles. A number of studies have been dedicated to the investigations of the heat effects caused by the surface enthalpy of nanoparticles. 13−18 However, most of these describe the heat effects for ceramic nanoparticles and/or are theoretical estimates. 14,16 −18 The purpose of the present work is to provide experimental heat effect data for the addition of Co in bulk and nanosized form into the liquid SAC387 alloy using drop calorimetry. The integral and partial enthalpies of mixing for the quaternary Ag−Co−Cu-Sn system in the Sn-rich corner are estimated. At the same time, calculations are performed to predict the expected differences in the data obtained for bulk and nanosized Co caused by the surface enthalpy of Co nanoparticles.
EXPERIMENTAL DETAILS
The calorimetric measurements were carried out using a Calvet-type twin microcalorimeter system, based on a commercial wire wound resistance furnace (HTC-1000, SETARAM, Lyon, France) having two thermopiles with more than 200 thermocouples, equipped with an self-made automatic drop device for up to 30 drops; control and data evaluation were performed with Lab View and HiQ. This system was described in detail by Flandorfer et al. 19 The measurements were performed in BN crucibles under Ar flow (99.999 vol %, purification from oxygen, approximately 30 mL/min). At the end of each series, the calorimeter was calibrated by five drops of standard α-Al 2 O 3 provided by NIST (National Institute of Standards and Technology, Gaithersburg, MD). The time interval between individual drops was usually 40 min, and the acquisition interval of the heat flow was 0.5 s. The obtained signals were recorded and integrated. The measured enthalpy (integrated heat flow at constant pressure) is where n i is the number of moles of the added element i, H m denotes molar enthalpies, T D is the drop temperature, ΔH Signal * is the measured enthalpy in J·mol −1 , and T M is the calorimeter temperature of the respective measurement in Kelvin. The molar enthalpy difference (H m,i,T m − H m,i,T D ) was calculated using the SGTE data for pure elements. 20 Because of the rather small masses of added component i, partial enthalpies can be calculated directly as The integral molar enthalpy of mixing, Δ mix H, was calculated by summing the respective reaction enthalpies and division by the total molar amount of substance, where n j stands for the molar amount of substance which was already present in the crucible: Pure metals of high purity (99.99%, Alfa Aesar, Karlsruhe, Germany) were used without further purification as well as commercial nanosized Co (99.9%, average particle size 28 nm, IoLiTec Nanomaterials, Heilbronn, Germany). According to the technical data sheet, 21 the BET surface area of the nanosized Co particles was about A BET = (50 ± 10)·10 3 m 2 ·kg −1 . The SAC387 alloys were prepared from pure components sealed in quartz ampules and kept in the furnace at 1173 K for 2 weeks. All operations with nano Co were performed in a glovebox (M.Braun, LabMaster 130) in an atmosphere of purified Ar (O 2 and H 2 O < 5 ppm each). The calorimetric measurements were carried out by the addition of solid Co in bulk and nanosized form into liquid SAC387 alloys at four different temperatures from 673 to 1173 K. In the second case, Co nanoparticles were first packed into a SAC387 foil with a thickness of about 50 μm which had been formed using a foil rolling mill. The measurements with additions of packed nano Co were started by dropping five pieces of SAC387 foil in order to determine and, subsequently, subtract the heat effect of the The Journal of Physical Chemistry C Article SAC387 foil from the obtained measured enthalpy. To prove the accuracy of this procedure, we also performed a few measurements by packing bulk Co into the SAC387 foil. For instance, the second measurement runs at 1073 and 873 K were performed in such a way. The starting values of Δ mix H for the ternary Ag−Cu−Sn subsystem required for the evaluation of the integral molar enthalpy of mixing for quaternary liquid Ag−Co−Cu-Sn alloys were calculated by a Redlich−Kister− Muggianu polynomial using experimental data taken from Luef et al. 22 Random as well as systematic errors of drop calorimetry depend on the calorimeter construction, calibration procedure, signal integration, and "chemical errors", for example, incomplete reaction or impurities. Considering many calibration measurements done by dropping NIST standard sapphire, the standard deviation can be estimated to be less than ±1%. The systematic errors are mainly caused by parasitic heat flows, baseline problems at signal integration, and dropping and mixing problems. One can estimate that the random error of the measured enthalpy is ±150 J.
Selected furnace-cooled alloys after calorimetric runs were investigated by scanning electron microscopy (SEM) and X-ray diffraction (XRD) to check for complete dissolution of the dropped component. The powder XRD measurements were done on a Bruker D8 diffractometer at ambient temperature using Ni-filtered Cu Kα radiation (accelerating voltage 40 kV, electron current 40 mA). The diffractometer operates in the θ/2θ mode. The powder was fixed with petroleum jelly on a single crystal silicon sample carrier which was rotated during the measurement. The detection unit was a Lynxeye strip The electron microscope Zeiss Supra 55 VP was used for metallographic investigations. The excitation energy of the electron beam was 15−20 kV; backscattered electrons were detected in order to visualize the surfaces of the samples. The chemical analyses of the sample phases were performed using the energy dispersive X-ray (EDX) technique with the two characteristic spectral lines of Cu (K-line) and Sn (L-line).
Standard deviations for the chemical compositions obtained from EDX were about ±1 at%.
RESULTS AND DISCUSSION
The molar enthalpy data for the additions of bulk Co into the liquid SAC387 alloy are presented in Tables 1A−1G. Since the Co additions in this paper are presented in at. %, the composition of the SAC387 master alloy is also given in at. % (Sn-4.1Ag-1.3Cu) below. A kink in the concentration dependencies of the integral enthalpies of mixing as well as constant partial enthalpy values indicate the transition of the investigated The Journal of Physical Chemistry C Article quaternary system from the liquid to the semisolid state; that is, it indicates the beginning of the precipitation of a solid phase ( Figure 1). Thus, the constant values of the partial enthalpy of mixing beyond 4 at. % Co in Figure 1 mark the liquidus boundary at 873 K. Based on the Co concentration dependence of the partial enthalpy of mixing, the respective liquidus limits were estimated at different temperatures. All values with the bold font in Tables 1A−1G refer to compositions outside the single-phase liquid state. It should also be noted that the obtained molar enthalpy data for quaternary Ag−Co−Cu−Sn alloys are practically identical to our recent data for the binary Co−Sn system. 23 The difference between the partial enthalpies of mixing for the addition of bulk Co to liquid Sn and to the liquid Sn-4.1Ag-1.3Cu alloy does not exceed 1 kJ·mol −1 (cf. Figure 1). This is not surprising because of the high content of Sn in the SAC alloy.
It should be noted that in our measurements at 873 K, we observed an additional exothermic reaction immediately after the main endothermic reaction, whereas such effects were not observed at other temperatures. The total reaction time did not exceed 1500 s.
The resulting heat effects obtained for the addition of nanosized Co into the liquid Sn-4.1Ag-1.3Cu alloy show a marked difference compared to the values for bulk Co additions (Table 2A−2E). This difference is concentration independent at the investigated temperatures and is equal to (−7.5 ± 1.0) × 10 3 J·mol −1 (cf. Figure 2a−c). It should be also noted that the average error of estimated data for ΔH Signal * and Δ mix H̅ Co does not exceed 1000 J·mol −1 .
The measured enthalpy (ΔH Signal ) consists generally of two terms (see eq 1); however, we think that the additional heat effect resulting in less positive values of ΔH Signal relates only to the first term of eq 1. The enthalpy of reaction (Δ r H) corresponds to the heat effects for interactions between atoms of the added component, i.e. liquid Co (remember that the standard state in Tables 1A−1G and Table 2A−2E is metastable liquid Co), and the liquid Sn-4.1Ag-1.3Cu alloy.
Therefore, this term should be the same independently whether Co is added in bulk or in nanoform. We suggest that this difference is caused by the excess enthalpy of the Co where ΔH i,nano ex is the excess surface enthalpy of Co nanoparticles in J·mol −1 , which has a positive value. This additional term should be connected with the decrease in the melting temperature and latent heat of nanosized Co particles similarly to other metals in nanosized form. 15,18,24 It is provided here with a negative sign (similar to H m,i,T D ), as the surface of the nanoparticles exists in these experiments only in the initial state.
It should be also noted that Table 2A−2E shows the recalculated molar enthalpies values including the excess enthalpy of Co nanoparticles term.
The partial and integral enthalpies of mixing, after taking into account the surface effect, are in good agreement. This is shown in Figure 3 where, as an example, the partial (Figure 3a) and integral enthalpies of mixing (Figure 3b) for additions of Co in bulk and nanosized form at 1073 K are plotted as a function of the Co content.
Crossing the liquidus line into a two-phase field is usually indicated by a kink in the integral enthalpy of mixing and by constant values in the partial mixing enthalpy. The relatively small difference in the heat effects for minor additions of Co The Journal of Physical Chemistry C Article both in the liquid and the semisolid Ag−Co−Cu−Sn alloys allows the estimation of the concentration of this transition only from the partial enthalpy of mixing data. These difficulties for the liquidus limit estimation were already pointed out for the investigation of the enthalpies of mixing in the binary Co−Sn system. 23 The estimated limiting liquidus concentration values in the present study, approximately 2 at. % Co at 673 K, 4 at. % Co at 873 K, and 14 at. % Co at 1073 K are slightly larger (up to 1−2 at. %) than those for the binary Co−Sn system. 25,26 However, it is still suggested that these transitions are connected with the precipitation of CoSn 2 (at 673 K) and CoSn (at 873 and 1073 K), in analogy to the corresponding binary phase diagram. 25 In order to prove that all pieces of the solid component dropped into the liquid bath had completely dissolved, selected alloys were investigated by means of SEM-EDX and powder XRD measurements after the calorimeter had cooled. The results of phase analyses along with BSE images of two exemplary alloys can be found in Table 3. No residual pure Co was found in the investigated samples. However, even after slow cooling in the calorimeter, the samples are not in an equilibrium state. This is obvious by the presence of six different phases in a four-component system. Nevertheless, the absence of (Co) indicates full mixing and precipitation of Co-poor phases, either during measurement (beyond the liquidus limit) or during cooling after the measurement. The Journal of Physical Chemistry C
Article
The XRD phase analysis fully confirmed the phases that had been found by SEM-EDX. Based on the SEM-EDX results Cu atoms replaced Co in the CoSn 3 compound, which was formed on cooling at 1226 ± 2 K. 26
THEORETICAL CONSIDERATIONS
In the present work, it is estimated that the term relating to the excess enthalpy of nanosized Co is practically the same for all investigated temperatures and equals to about (7.5 ± 1.0)·10 3 J·mol −1 . This term relates to the surface enthalpy of 1 mol of nanosized Co as where ΔH i,surf is the surface enthalpy in J m −2 and A s is the surface area for 1 mol of nanoparticles in the unit of m 2 mol −1 . Assuming a strictly spherical shape of the nanoparticles, A s can be expressed as where V M is the molar volume; M is the molar mass; ρ is the density; and r is the radius of the particles. Inserting the corresponding values for Co, i.e. M = 58.933 × 10 −3 kg·mol −1 and ρ = 8.890·10 3 kg·m −3 , and the radius of the Co nanoparticles (∼14 nm) results in a molar surface area of about 1.42 × 10 3 m 2 ·mol −1 . As mentioned above, the BET surface area of the Co nanopowder used here is equal to A BET = (50 ± 10)·10 3 m 2 ·kg −1 corresponding to A s = (2.95 ± 0.59)·10 3 m 2 ·mol −1 . 21 The obtained discrepancy between calculated and technical values is most probably caused by the variation in size (particle size range is given as 0−60 nm, with an average size of 28 nm) and shape of the employed nanoparticles; these values, in turn, combined with the experimental value of (7.5 ± 1.0)·10 3 J·mol −1 , give a surface enthalpy of about (2.85 ± 0.55) J·m −2 .
On the basis of the data presented above, it was decided to describe the observed phenomenon, related to the nano Co additions, according to the thermodynamic properties of nanosized particles. The molar enthalpy of nanoparticles can be expressed as 27 where H s, nano (J·mol −1 ) is the molar enthalpy of the solid nanoparticles, H s, bulk (J·mol −1 ) is the molar enthalpy of the bulk solid; A spec (m −1 ) is the specific surface area of the nanoparticles defined as the ratio of their surface area to their volume, σ sg,H,T D (J·m −2 ) is the enthalpy term of the surface energy of the solid nanoparticles at the dropping temperature. Although both molar volume and surface energy are T-dependent quantities, the total value of ΔH i,nano ex (as it is at the dropping temperature) is lost due to the elimination of the surface of the nanoparticles upon their dissolution in the liquid alloy, and therefore, the measured nanoheat-effect is not | 3,798.2 | 2016-01-20T00:00:00.000 | [
"Engineering",
"Materials Science"
] |
Thermal Transport on Graphene-Based Thin Films Prepared by Chemical Exfoliations from Carbon Nanotubes and Graphite Powders
: Thermal conductivities ( k ) of different graphene nanosheet (GN)-based heat sinks are investigated within the temperature range of 323–423 K. One- and two-step modified Hummers’ methods are adopted to chemically exfoliate GNs from two kinds of carbon precursors: carbon nanotubes (CNTs) and graphite powders. The two-step method offers an improved exfoliation level of GN products, especially for the CNT precursor. Experimental results show that the GN-based heat sink—exfoliated from graphite powders after the two-step approach—delivers an enhanced k value to 2507 W/m K at 323 K, as compared to the others. The k value is found to be a decreasing function of the porosity of the heat sink, revealing the importance of solid/void fraction (i.e., volumetric heat capacity). The improved thermal efficiency mainly originates from the long phonon mean free path and the low void fraction of GN-based heat sinks, thus inducing highly efficient thermal transport in the GN framework.
Introduction
Recently, one-and two-dimensional carbon nanostructures-especially carbon nanotubes (CNTs) and graphene nanosheets (GNs)-have attracted a great deal of scientific and technological attention, owing to their extraordinary electrical and thermal properties, exhibiting their applicable feasibility [1,2]. Both CNTs and GNs have been demonstrated to display high theoretical thermal conductivities (k), thus heralding beneficial applications in heat exchangers and thermal management systems. It is generally recognized that low-temperature operation and efficient heat dissipation are key factors in determining the lifetime and the speed in current electronic and photonic devices [1,3]. Industrially, the most commonly-used materials for heat sinks are still Al and Cu, which possess lower k values and higher true densities, as compared to carbonaceous materials. Hence, there is torpid progress on the thermal management of interfacial solid surfaces, thus hampering the miniaturization of electronic and photonic devices [4]. To resolve the above problems, one strategy is to adopt nanostructural carbon materials (e.g., GNs) as heat sinks, capable of conducting heat efficiently and thus preventing structural damage to electronic components [5].
It has been reported that single-layer GNs possess a high theoretical k value of 4840-5300 W/m K [6][7][8][9], motivating many scientists and researchers to consider GN-based heat sinks for nanoscale device application. However, an obvious gap exists between the actual k values of GN-based layers or heat sinks and the theoretical ones. As to the nano-structured carbons, the k values reported thus far include aligned functionalized multilayer GN architecture (112 W/m K at 298 K) [1], three-dimensional CNT/GN framework (1991 W/m K at 323 K) [2], graphene nanoribbons (2300 W/m K at 300 K) [7], GN-Cu-GN heterogeneous films (373 W/m K at 300 K) [10], and N-doped GN/Cu composite film (543 W/m K) [11]. The major gap can be attributed to the fact that the efficiency of thermal transport is mainly influenced by the presence of topological defects, surface roughness, surface heterogeneity, and porosity in carbon-based heat sinks [7]. Among the properties of carbon-based materials, the porosity may play a crucial role in affecting the thermal conduction in the heat sinks if the pores are full of air, since air has a very low k value (~0.024 W/m K at ambient temperature). However, few reports have discussed the relationship between the porosity and the k values in GN-based heat sinks.
Accordingly, this work aims to explore the influence of porosity on the efficiency of thermal transport in GNs. An efficient Hummers' method was adopted to chemically exfoliate graphene oxide (GO) sheets from two kinds of carbon precursors (i.e., CNTs and graphite powders). To ensure the completion of chemical exfoliation, the modified Hummers' approach was carried out twice, and the as-prepared GO sheets were thermally reduced in an H 2 -containing atmosphere. The thermal transport properties of GN-based heat sinks were systematically investigated within a wide temperature range (323-423 K), demonstrating the feasibility of GN-based heat sinks for practical applications.
Materials and Methods
Compared to the original method of Hummers [12], this work made some modifications in oxidizing agent concentration, oxidation period, and ratio of precursor to oxidation agent to prepare GO sheets, as reported previously [13,14]. Two types of carbon precursors-CNTs (AzTrong Inc., Hsinchu, Taiwan) and graphite powders (Taiwan Maxwave Co., Taoyuan, Taiwan)-were of commercial origin. The multi-layered CNT product was grown by chemical vapor deposition method, using Ni-based nanoparticles and ethylene as catalyst and carbon source, respectively. The graphite powder (purity: 99.5%, average particle size: 3 µm) employed here widely serves as a conducting agent for Li-ion batteries. The one-step method of Hummers for producing GO sheets can be briefly described in the following four steps. In step (i), 5 g carbon powders (i.e., CNTs and graphite powders) were put into 115 mL concentrated H 2 SO 4 and well dispersed in an ice bath for 1.5 h. Afterwards, 15 g KMnO 4 was slowly added into the slurry within 1 h while keeping the temperature less than 5 • C. In step (ii), the slurry was stirred by using a magnetic bar and then heated to 35 • C. Next, 200 mL distilled water was mixed well with the slurry. The temperature of the slurry was gradually increased to 95 • C and then kept at that temperature for 0.5 h. The purpose of pouring distilled water into the carbon slurry is to dilute the sulfuric acid, preventing excessive gas generated from the strong chemical reaction. We aimed to prevent a dramatic increase in operating temperature, avoiding the hazard potential of sudden boiling. Moreover, fast exfoliation induced a re-stacking of graphene sheets without diluting the strong oxidizers, based on our in-house study. In step (iii), the chemical oxidation was terminated by adding 500 mL of H 2 O 2 solution. The addition of H 2 O 2 was able to remove the residual potassium permanganate and manganese dioxide, leading to an increase in reactive area for the implantation of surface functionalities during the chemical oxidation process. In step (iv), one thermal reduction process was conducted at 400 • C under 5 vol % H 2 atmosphere in a horizontal furnace. As for the two-step Hummers' method, step (ii) was repeated once again to reinforce the oxidation extent on the carbon samples (i.e., the second replacement of MnO 3 + ions).
The as-prepared GN nanostructures were observed by field-emission scanning electron microscopy (FE-SEM; JSM-6701F, JEOL, Kyoto, Japan) and transmission electron microscopy (TEM; JEM-2100, JEOL). An X-ray diffraction (XRD; Labx XRD-6000, Shimadzu, Kyoto, Japan) spectroscope equipped with Cu-Kα radiation emitter was used to characterize the crystallinity of GN samples. Fourier transformed infrared (FT-IR) spectroscopy was adopted to analyze oxygen functionalities of GN samples. The FT-IR spectra were obtained using a Nicolet Avatar 360 FT-IR spectrometer (SpectraLab Scientific Inc., Markham, ON, Canada). For each sample, 32 scans in the spectral ranges were recorded with a resolution of 4 cm −1 . The crystalline structure of GN powders was characterized by using Raman spectroscopy (Micro-Raman spectrometer, Renishaw, Gloucestershire, UK).
For the measurement of k values, the GN-based heat sinks could be prepared per the following description. First, a slurry-coating technique was adopted to prepare GN-based layers onto Cu foil. Herein, all GN samples were uniformly mixed with a binder (poly-vinylidenefluoride) with the weight ratio of 80:20 in N-methyl pyrrolidinone (NMP) solvent to form the graphene slurries. Second, the graphene slurries were well blended with a three-dimensional mixer by using zirconia balls for 15 min. The as-prepared slurries were carefully pasted on the foil substrates (thickness:~10 µm) with a doctor blade, followed by evaporating the solvent (i.e., NMP) with a blow dryer. The as-prepared graphene layers with an area of 2 × 5 cm 2 were then dried at 120 • C in a vacuum oven overnight. Finally, the electrode sheets were carefully pressed under a pressure of approximately 193.6 atm. The apparent density ( app ) of heat sinks was determined by the ratio of the mass to a given volume. For the determination, the heat sink was with an area of 2 × 5 cm 2 and its real thickness was measured by optical microscope. The app value was thus determined by using the ratio of the weight to the measured volume. For accuracy, three pieces of heat sinks were used to get their average readings. One test platform system for determining the k value of the heat sinks was reported in previous studies [2,4].
High-purity (i.e., 99.95%) Cu plates with different thicknesses served as the reference material to calibrate the accuracy of k values. The calibration curves of operating temperature against heat transfer from an electrical resistance heater could be obtained. The calibration temperatures ranged from 323 to 423 • C, appropriate for the thermal management system for chips in consumer electronics. To confirm the accuracy, Al foil was used as the other reference to reveal the small deviations (<2.0%) at different temperatures (e.g., the k value of Al: ca. 237 W/m K at ambient temperature).
All GN-based heat sinks were insulated with heat-preservation cotton to avoid any heat dissipation. The temperature reading was monitored and recorded by using an array of five thermocouples (K-type, accuracy: 0.1 • C): one was located at the heater, three at the heat sink, and one for measuring the ambient temperature. The apparent k values of the heat sink could be determined by using Fourier's law, based on one-dimensional heat conduction [15]. The k values were determined by comparing the temperature drops across the heat sink, followed by the calculations of Fourier's law. Herein, each reading was averaged from the k values at different locations of heat sinks.
Results and Discussion
The as-prepared GN samples were designated as two series: GN-CNT1 and GN-CNT2 (precursor: CNT, 1 and 2: one-and two-step chemical exfoliation) and GN-G1 and GN-G2 (precursor: graphite powders, 1 and 2: one-and two-step chemical exfoliation). Figure 1a,b show FE-SEM images of pristine CNTs and graphite powders, respectively. It can be seen that the CNTs look like coiled-type tubes with an average diameter of 30-50 nm, while the pristine graphite powders show a typical layer-stacking morphology. High-resolution transmission electron microscopy (HR-TEM) micrographs for GN samples prepared by the one-and two-step Hummers' method from CNTs and graphite powders are illustrated in Figure 1c-f, showing different morphologies. Apparently, the one-step Hummers' route cannot completely exfoliate GNs from CNTs and graphite powders. As observed from Figure 1c,d, both pristine CNTs and graphite powders basically keep their original shapes (i.e., tubular and layer-stacking shape). Regarding the two-step approach, the GN-CNT2 sample looks like a broken graphene layer, in which the tubular-type structure was unzipped through the two-step exfoliation route. The GN-G2 sample seems like a transparent and soft silk with few layers, resulting from the full chemical exfoliation. Based on the observation, the two-step chemical procedure is capable of improving the exfoliation level of GN samples for both carbon precursors.
XRD was adopted to further characterize the crystalline structure of GN samples. Figure 2 depicts typical XRD patterns of GN samples, in which both pristine CNTs and graphite powders possess high crystallinity of graphite lattices. An obvious (002) diffraction peak at 26.2-26.4 • can be seen; i.e., the d 002 -interlayer spacing distance: 3.41 Å (CNTs) and 3.36 Å (graphite powders). The (002) diffraction peaks tend to shift and broaden after the one-and two-step chemical exfoliation process, demonstrating the presence of GNs [4]. This reflects an obvious increase in the interlayer distance between each graphene layer. Compared to the original values, the d 002 -interlayer spacing distances are up to 3.82 Å (GN-G1), 3.85 Å (GN-G2), 3.43 Å (GN-CNT1), and 3.46 Å (GN-CNT2). This finding implies that the CNT precursor shows better chemical resistance than graphite powder. This possibly originates from one-dimensional tubular structure, in which chemical oxidants (e.g., SO 4 2− ions) tend to attack and then oxidize both opened ends of CNT during the chemical exfoliation process. To sum up, the GNs prepared from graphite powders can achieve an efficient exfoliation level when compared to the GNs from CNT precursor. . This finding implies that the CNT precursor shows better chemical resistance than graphite powder. This possibly originates from one-dimensional tubular structure, in which chemical oxidants (e.g., SO4 2− ions) tend to attack and then oxidize both opened ends of CNT during the chemical exfoliation process. To sum up, the GNs prepared from graphite powders can achieve an efficient exfoliation level when compared to the GNs from CNT precursor. FT-IR spectroscopy was adopted to analyze the distribution of oxygen functional groups on different GN samples, as shown in the Supporting Information (see Figure S1a). The FT-IR spectra reveal the presence of C=O and C-O groups centered at their specific wavenumbers [16]. It is worth noting that both GN-G1 and GN-CNT1 samples exhibit stronger intensity of transmittance bands compared to the others. This reflects that the first-step chemical oxidation efficiently facilitates the oxidation level of graphene samples, whereas the second-step oxidation-followed by thermal reduction-is capable of stripping oxygen functionalities from basal plane or the edge of graphene samples. Thus, both GN-G2 and GN-CNT2 samples possess a low oxidation extent. As shown in Figure S1b, typical Raman spectra show two characteristic peaks occurring at 1350 cm −1 and 1580 cm −1 , corresponding to the D and G bands, respectively. The intensity ratio of D to G band (i.e., ID/IG) serves as a key factor in determining the graphitization degree of carbons, in which the D and G bands indicate the defects or structural disorder and the stacking of the graphite hexagon network plane [17,18]. As compared to the treated GN samples, the pristine G and CNT samples have lower ID/IG ratios-0.33 and 0.51, respectively. The ID/IG ratios of GN-G1 and GN-G2 samples are approximately 1.18-1.21, whereas the ID/IG ratios for both GN-CNT1 and GN-CNT2 samples are up to 1.29-1.45. This increased ID/IG ratio can be attributed to the presence of lattice distortion (e.g., sp 3 /sp 2 ratio) and surface functionalities (e.g., C-OOH and C=O groups) that are decorated in the stacking order of graphite lattice, presumably due to the residual oxygen functionalities after thermal reduction process. FT-IR spectroscopy was adopted to analyze the distribution of oxygen functional groups on different GN samples, as shown in the Supporting Information (see Figure S1a). The FT-IR spectra reveal the presence of C=O and C-O groups centered at their specific wavenumbers [16]. It is worth noting that both GN-G1 and GN-CNT1 samples exhibit stronger intensity of transmittance bands compared to the others. This reflects that the first-step chemical oxidation efficiently facilitates the oxidation level of graphene samples, whereas the second-step oxidation-followed by thermal reduction-is capable of stripping oxygen functionalities from basal plane or the edge of graphene samples. Thus, both GN-G2 and GN-CNT2 samples possess a low oxidation extent. As shown in Figure S1b, typical Raman spectra show two characteristic peaks occurring at 1350 cm −1 and 1580 cm −1 , corresponding to the D and G bands, respectively. The intensity ratio of D to G band (i.e., I D /I G ) serves as a key factor in determining the graphitization degree of carbons, in which the D and G bands indicate the defects or structural disorder and the stacking of the graphite hexagon network plane [17,18]. As compared to the treated GN samples, the pristine G and CNT samples have lower I D /I G ratios-0.33 and 0.51, respectively. The I D /I G ratios of GN-G1 and GN-G2 samples are approximately 1.18-1.21, whereas the I D /I G ratios for both GN-CNT1 and GN-CNT2 samples are up to 1.29-1.45. This increased I D /I G ratio can be attributed to the presence of lattice distortion (e.g., sp 3 /sp 2 ratio) and surface functionalities (e.g., C-OOH and C=O groups) that are decorated in the stacking order of graphite lattice, presumably due to the residual oxygen functionalities after thermal reduction process. The variation of k values of heat sinks with operating temperature (T) is illustrated in Figure 3. This figure clearly indicates that the k value shows a decreasing trend with temperature, revealing that the thermal transport in GN-based heat sink is mainly dominated by phonon-boundary scattering. This is because the relationship between k value and T is inconsistent with 1/T temperature dependence within the temperature range, referring to Umklapp phonon-phonon scattering [19,20]. The relation delivers two crucial messages: (i) the GN-based heat sinks prepared from graphite powders deliver better thermal diffusivity than that from CNT precursor, and (ii) the two-step exfoliation route significantly improves the efficiency of thermal transport. The experimental results demonstrate that the GN-G2 sample showed the highest k values as compared to the others, reaching as high as 2,507 W/m K at 323 K. In principle, the k value is proportional to a product, Cpvl, where Cp is the heat capacity per unit of volume, v the speed of sound, and l the phonon mean free path [5]. The magnitude of Cp is strongly related to the solid/void fraction (i.e., porosity) in heat sink, whereas the l value depends on the phonon scattering from carbon boundaries (i.e., grain size), point defects, and Umklapp processes. A pioneering study has pointed out that the CNT-based heat sink consists of highly-resistive thermal junctions between nanotubes [9], thus, leading to smaller l value than the other heat sinks. As observed from HR-TEM and XRD analyses, the CNTs still maintain onedimensional tubular structure after the one-step exfoliation route. This could be why the GN-CNT1 heat sink displays the lowest thermal transport capability among the heat sinks. With increasing the extent of exfoliation (i.e., after the two-step exfoliation route), the presence of GNs in the heat sinks beneficially facilitates the heat transport, which has been proved by the GNs with high intrinsic k value. Without any thermal resistive junctions, the k value of two-dimensional nanostructures is primarily a function of large mean free path across the strong sp 2 bonds in the networks of carbon atoms [21]. Therefore, the stacking of GN architecture can be taken into account as a conductive framework, offering more thermal diffusion paths for heat dissipation.
To inspect the effect of solid-void fraction, the porosity of GN-based heat sinks was measured by using the formula: (1 − ρapp/ρt) × 100% [22], where ρapp and ρt represent apparent and true density of the heat sinks, respectively. The estimated porosity of GN-based heat sinks has an order as follows: GN-CNT1 (35.1%) > GN-CNT2 (22.0%) > GN-G1 (20.7%) > GN-G2 (19.9%). It is worth noting that the GN-CNT1 heat sink is mainly composed of coiled-type tubular structure, allowing a large amount of air to be trapped in the nanostructure. The air pockets in the heat sink would raise the porosity and then lower both the Cp and l values, which is unfavorable for the heat transport in GNs. The variation of k values of heat sinks with operating temperature (T) is illustrated in Figure 3. This figure clearly indicates that the k value shows a decreasing trend with temperature, revealing that the thermal transport in GN-based heat sink is mainly dominated by phonon-boundary scattering. This is because the relationship between k value and T is inconsistent with 1/T temperature dependence within the temperature range, referring to Umklapp phonon-phonon scattering [19,20]. The relation delivers two crucial messages: (i) the GN-based heat sinks prepared from graphite powders deliver better thermal diffusivity than that from CNT precursor, and (ii) the two-step exfoliation route significantly improves the efficiency of thermal transport. The experimental results demonstrate that the GN-G2 sample showed the highest k values as compared to the others, reaching as high as 2,507 W/m K at 323 K. In principle, the k value is proportional to a product, C p vl, where C p is the heat capacity per unit of volume, v the speed of sound, and l the phonon mean free path [5]. The magnitude of C p is strongly related to the solid/void fraction (i.e., porosity) in heat sink, whereas the l value depends on the phonon scattering from carbon boundaries (i.e., grain size), point defects, and Umklapp processes. A pioneering study has pointed out that the CNT-based heat sink consists of highly-resistive thermal junctions between nanotubes [9], thus, leading to smaller l value than the other heat sinks. As observed from HR-TEM and XRD analyses, the CNTs still maintain one-dimensional tubular structure after the one-step exfoliation route. This could be why the GN-CNT1 heat sink displays the lowest thermal transport capability among the heat sinks. With increasing the extent of exfoliation (i.e., after the two-step exfoliation route), the presence of GNs in the heat sinks beneficially facilitates the heat transport, which has been proved by the GNs with high intrinsic k value. Without any thermal resistive junctions, the k value of two-dimensional nanostructures is primarily a function of large mean free path across the strong sp 2 bonds in the networks of carbon atoms [21]. Therefore, the stacking of GN architecture can be taken into account as a conductive framework, offering more thermal diffusion paths for heat dissipation.
To inspect the effect of solid-void fraction, the porosity of GN-based heat sinks was measured by using the formula: (1 − app / t ) × 100% [22], where app and t represent apparent and true density of the heat sinks, respectively. The estimated porosity of GN-based heat sinks has an order as follows: GN-CNT1 (35.1%) > GN-CNT2 (22.0%) > GN-G1 (20.7%) > GN-G2 (19.9%). It is worth noting that the GN-CNT1 heat sink is mainly composed of coiled-type tubular structure, allowing a large amount of air to be trapped in the nanostructure. The air pockets in the heat sink would raise the porosity and then lower both the C p and l values, which is unfavorable for the heat transport in GNs. After the two-step exfoliation process, the porosity of GN-based heat sinks tends to be reduced for both carbon precursors. Figure 4a-d shows cross-sectional FE-SEM images of different GN-based heat sinks. As expected, the GN-CNT1 sample possesses a number of voids and cavities, possibly leading to the formation of air pockets in the heat sink. In contrast, the GN-G2 sample is found to have a dense surface, not allowing air storage. The result of porosity analysis reflects that the exfoliation level on the GNs would effectively improve the density of GN stacking layers, thus reducing the porosity of heat sinks. The relationship between the porosity and the thermal conduction efficiency is depicted in Figure 5. As seen in Figure 5a, the GN-CNT1 heat sink displays the largest porosity but the lowest k value among the samples. The k value is illustrated in Figure 5b as a decreasing function of the The relationship between the porosity and the thermal conduction efficiency is depicted in Figure 5. As seen in Figure 5a, the GN-CNT1 heat sink displays the largest porosity but the lowest k value among the samples. The k value is illustrated in Figure 5b as a decreasing function of the porosity of GN-based heat sinks. This result reveals that the porosity of heat sinks, is strongly correlated to the volumetric C p value, playing an important role in affecting thermal diffusivity. The higher the surface density of GN-based heat sinks, the higher the thermal transport efficiency that can be achieved. However, it is worth noting that the k value is not a linear function of porosity. The improved thermal efficiency mainly originates from high l value and low void fraction of GN-based heat sinks, thus leading to highly efficient thermal transport in the carbon framework. This finding reveals that the efficiency of thermal transport of heat sinks depends not only on the intrinsic thermal properties (e.g., intrinsic k value), but also on the physical and mechanical preparation method (e.g., compressive treatment). Using the same test platform, the GN-G2 sample was found to possess higher k value than the other designs of heat sinks such as CNT/GN hybrid (1991 W/m K at 323 K) [2] and Al 2 O 3 -coated graphite (1128 W/m K at 323 K) [3]. It can be concluded that the selection of carbon materials is the major contributor to the thermal conduction. However, the preparation of heat sinks including chemical composition, rolling process, and compression ratio also is also crucial in affecting the thermal diffusivity. The resultant heat sink used in the present work shows a high compression ratio that may alleviate the effect of O/C ratio on the thermal conduction performance. On the basis of experimental results, the GNs could serve as a feasible material for high-performance thermal devices, such as heat exchangers and heat sinks for consumer electronic devices or integrated circuit chips. Moreover, the k value of the GN-G2 sample is approximately 6.5 and 10.5 times higher than that of Cu and Al foils, respectively. This result demonstrates that the robust design of the GN-G2 sample is superior to that of conventional metal-based heat sinks due to its lightweight, flexibility, and excellent thermal conduction. The relationship between the porosity and the thermal conduction efficiency is depicted in Figure 5. As seen in Figure 5a, the GN-CNT1 heat sink displays the largest porosity but the lowest k value among the samples. The k value is illustrated in Figure 5b as a decreasing function of the porosity of GN-based heat sinks. This result reveals that the porosity of heat sinks, is strongly correlated to the volumetric Cp value, playing an important role in affecting thermal diffusivity. The higher the surface density of GN-based heat sinks, the higher the thermal transport efficiency that can be achieved. However, it is worth noting that the k value is not a linear function of porosity. The improved thermal efficiency mainly originates from high l value and low void fraction of GN-based heat sinks, thus leading to highly efficient thermal transport in the carbon framework. This finding reveals that the efficiency of thermal transport of heat sinks depends not only on the intrinsic thermal properties (e.g., intrinsic k value), but also on the physical and mechanical preparation method (e.g., compressive treatment). Using the same test platform, the GN-G2 sample was found to possess higher k value than the other designs of heat sinks such as CNT/GN hybrid (1,991 W/m K at 323 K) [2] and Al2O3-coated graphite (1,128 W/m K at 323 K) [3]. It can be concluded that the selection of carbon materials is the major contributor to the thermal conduction. However, the preparation of heat sinks including chemical composition, rolling process, and compression ratio also is also crucial in affecting the thermal diffusivity. The resultant heat sink used in the present work shows a high compression ratio that may alleviate the effect of O/C ratio on the thermal conduction performance. On the basis of experimental results, the GNs could serve as a feasible material for high-performance thermal devices, such as heat exchangers and heat sinks for consumer electronic devices or integrated circuit chips. Moreover, the k value of the GN-G2 sample is approximately 6.5 and 10.5 times higher than that of Cu and Al foils, respectively. This result demonstrates that the robust design of the GN-G2 sample is superior to that of conventional metal-based heat sinks due to its lightweight, flexibility, and excellent thermal conduction.
Conclusions
We have presented the efficiency of thermal transport in GN-based heat sinks, chemically exfoliated from CNTs and graphite powders using one-and two-step modified Hummers' method, within the temperature region of 323-423 K. The CNT precursor showed strong chemical resistance to chemical oxidizers, whereas the graphite powders could be chemically exfoliated to GN products with high exfoliation extent. The two-step method displayed an improved exfoliation level of GN products-especially for CNT precursor. The GN-based heat sink, exfoliated from graphite powders after the two-step approach could attain the highest k value of 2507 W/m K at 323 K, as compared to
Conclusions
We have presented the efficiency of thermal transport in GN-based heat sinks, chemically exfoliated from CNTs and graphite powders using one-and two-step modified Hummers' method, within the temperature region of 323-423 K. The CNT precursor showed strong chemical resistance to chemical oxidizers, whereas the graphite powders could be chemically exfoliated to GN products with high exfoliation extent. The two-step method displayed an improved exfoliation level of GN products-especially for CNT precursor. The GN-based heat sink, exfoliated from graphite powders after the two-step approach could attain the highest k value of 2507 W/m K at 323 K, as compared to the others. The k value as a decreasing function of the porosity of GN-based heat sinks was explored. This result reflected that the porosity of heat sink-strongly correlated to the volumetric C p value-is one of the crucial factors in determining the thermal diffusivity. The lower the porosity of GN-based heat sinks, the higher the thermal transport efficiency that can be achieved. Since the k value is not a linear function of porosity, its improved value could be attributed to both high l value and low void fraction of GN-based heat sinks, thus leading to highly efficient thermal transport in the carbon framework. As a result, this work delivered a feasible possibility of GN-based films for the future development of heat exchanger and heat sink applications.
Supplementary Materials: The following are available online at http://www.mdpi.com/2079-6412/7/9/138/s1. Figure S1: (a) FT-IR patterns and (b) Raman spectra of different graphene samples. This reflects that the first-step chemical oxidation efficiently facilitates the oxidation level of graphene samples, whereas the second-step oxidation, followed by thermal reduction, is capable of stripping oxygen functionalities from basal plane or edge of graphene samples. | 6,862.8 | 2017-09-03T00:00:00.000 | [
"Materials Science",
"Engineering",
"Physics",
"Chemistry"
] |
Applications of Photonics in Agriculture Sector: A Review
The agricultural industry has made a tremendous contribution to the foundations of civilization. Basic essentials such as food, beverages, clothes and domestic materials are enriched by the agricultural industry. However, the traditional method in agriculture cultivation is labor-intensive and inadequate to meet the accelerating nature of human demands. This scenario raises the need to explore state-of-the-art crop cultivation and harvesting technologies. In this regard, optics and photonics technologies have proven to be effective solutions. This paper aims to present a comprehensive review of three photonic techniques, namely imaging, spectroscopy and spectral imaging, in a comparative manner for agriculture applications. Essentially, the spectral imaging technique is a robust solution which combines the benefits of both imaging and spectroscopy but faces the risk of underutilization. This review also comprehends the practicality of all three techniques by presenting existing examples in agricultural applications. Furthermore, the potential of these techniques is reviewed and critiqued by looking into agricultural activities involving palm oil, rubber, and agro-food crops. All the possible issues and challenges in implementing the photonic techniques in agriculture are given prominence with a few selective recommendations. The highlighted insights in this review will hopefully lead to an increased effort in the development of photonics applications for the future agricultural industry.
Introduction
Light constitutes a collection of particles known as photons, propagated in the form of waves [1]. In physics, light often relates to radiation in the entire electromagnetic spectrum, encompassing X-rays, ultraviolet, visible light, infrared, and microwaves among others [2]. The unique electromagnetic properties of light have intrigued academics across the globe and the earliest study can be traced back to the early 17th century [3]. As time passes, the accumulation of knowledge and technological advancement have gradually shaped the canvas for light-related research, leading to the establishment of the field of optics and photonics.
Optics can be defined as a branch of physics that studies the behavior and properties of light as well as the interaction of light with other matter [2]. Meanwhile, photonics can be regarded as the application of light through the systematic generation, control and detection of photons [2,4]. Despite the distinction between optics and photonics, both terminologies have often been used interchangeably in the literature to collectively represent the science and application of light [1].
Optics and photonics have influenced various engineering applications, transforming the landscape of various fields and improving the lives of mankind. One of the main applications of optics and photonics can be seen in the field of communications. Knowledge of optics and photonics has been used to develop optical fibers which help to cater for the needs of broadband Internet service in this "data hungry" era. Furthermore, optics and photonics have been used in the manufacturing of modern displays such as liquid crystal display (LCD), organic light-emitting diode (OLED), flexible display and such. Solar cells for energy harnessing too illustrate another application of optics and photonics. Not least, optics and photonics have also been applied in more sophisticated areas such as security surveillance, medical imaging, quantum computing and more [1].
Amidst the modern and complex solutions discussed earlier, it often slipped our minds that optics and photonics can be readily integrated into the field of agriculture. The simplest examples would be the adjustment of plantation direction for optimum sunlight exposure, as well as the usage of incandescent light bulbs in egg incubation and hatching [5]. Over recent decades, academics have been alerted to the potential of optics and photonics in the agricultural industry. This has led to progressive developments that utilize optics and photonic techniques in maximizing the quality and productivity of agricultural products.
This paper aims to review some of the most popular optics and photonic techniques in agriculture, namely imaging, spectroscopy and spectral imaging. In addition, existing applications of each technique in the agricultural industry will also be compiled. A comprehensive discussion will also be made to gauge the potential of exploiting optics and photonic techniques in the agricultural sector with the intention of improving the quality and productivity of the agricultural products at a reduced labor cost.
Classification of Photonics Systems in Agriculture
Quantity and quality have always been the primary foci in the field of agriculture. The governing of these attributes is anticipated to be more crucial in the upcoming years. This prediction is based on the constant increase in global population as well as heightened expectations for healthy food sources. However, the agricultural field faces great pressure under globalization. The transformation of the global economic landscape makes agricultural activities seem less profitable in contrast to other industrial activities. The outflow of the workforce makes it increasingly expensive and difficult to meet the demands of agricultural activities.
As a result, modern technology has been integrated into the agricultural field to maximize output efficiency at minimum labor force. Similar to other industries, automation systems have been applied in stages of agricultural activities to reduce a dependency on manual labor [6]. These systems require optics and photonics techniques to complement them, providing the required 'sight' for operations. These vision requirements have been fulfilled by optics and photonic techniques such as imaging, and spectral imaging. These techniques provide machine vision at high dynamic range, high resolution and high accuracy in a non-destructive, non-contact and robust manner [5]. In the subsections below, details of ESS configurations, their classifications and structures have been illustrated.
Imaging Technique
The imaging technique is analogous to the function of the human eye. It captures the image of the subject for necessary calculations and measurements before performing the final evaluations [7]. The imaging technique is essential for collecting spatial, color [6] and even thermal [8] information of the subject of interest. Therefore, imaging techniques are typically operated in an active manner. The active imaging technique involves image acquisition under two major light sources, namely visible light and infrared sources. Images under visible light can be easily acquired with any standard camera modules. On the other hand, images under exposure to infrared can be acquired with special infrared camera modules [8]. Image acquisition under visible light is similar to our daily photography. The image acquisition process under this light source is straightforward and images captured are usually rich in details and colors. However, complexity often arises while performing analysis on these images due to illumination variations. For instance, images captured outdoors vary under sunny and cloudy conditions. Meanwhile, images captured indoors is categorized by natural light, incandescent and fluorescent conditions [7].
The acquired image will then undergo pre-processing to convert it into an appropriate format before further analysis. Pre-processing tasks may include exposure correction, color balancing, noise reduction, sharpness increase or orientation change. Next, the process of feature detection and matching as well as segmentation is performed on the pre-processed image to extract the object or region of interest. Finally, the subject of interest is analyzed with proper analysis algorithms in the respective area of application [9].
The imaging technique can be easily applied in the simple analysis of static-positioned objects or even in more complex areas which involve moving targets, such as visual navigation and behavioral surveillance. These achievements were made possible by utilizing the spatial information acquired through the imaging technique for position triangulation and motion guidance [7,9]. In image processing, the computer imaging technique has been employed to create, edit, and display graphical images, characters, and objects. The computer image analysis technique is a broad field which consists of computer domains and applications in food quality evaluation [10,11], grading and the sorting of agricultural products [12,13], as well as harvesting the crops [14], and estimating moisture content in the drying stage for the storability of the food product [15]. Computer imaging contributes to the development of digital agriculture. For instance, weed detection and fruit grading systems with digital imaging techniques are cost effective systems in achieving ecological and economically sustainable agriculture [16].
Spectroscopy Technique
In contrast to the imaging technique, the spectroscopy technique enables the 'sight' of properties that are invisible to the naked eye. The spectroscopy technique functions by extracting spectral information from the sample of interest. The spectral information is obtained when light interacts with the composition of the sample. This interaction leads to changes in the intensity or frequency and wavelength of the initial light source, ultimately defining a spectrum which acts as the fingerprint of the sample [17].
Similar to the imaging technique, variations do exist for spectroscopy. These variations are categorized by the nature of interaction between the light source and the sample when the spectroscopy measurement is conducted. In the agricultural field, the commonly adopted spectroscopy techniques are ultraviolet-visible (UV-VIS) spectroscopy, fluorescence spectroscopy, infrared (IR) spectroscopy, and Raman spectroscopy [17].
Ultraviolet-Visible (UV-VIS) Spectroscopy
The ultraviolet-visible (UV-VIS) spectroscopy is conducted in both the ultraviolet (UV) and visible light (VIS) band, spanning wavelengths from 100 nm to 380 nm (UV) and from 380 nm to 750 nm (VIS). The principle governing the UV-VIS spectroscopy is Beer-Lambert's law, which is expressed by (1) and (2): ln where I 0 and I are intensity of light entering and leaving a sample respectively, ε is the extinction molar coefficient, c is the molar concentration of substance, l is the thickness of sample (cm), T is transmittance and A is absorbance [18]. A typical model that illustrates Beer-Lambert's law can be seen in Figure 1. It can be observed that as light propagates through a sample, a portion of the incidental light source will be absorbed by the molecules in the sample, while the remaining light rays will transmit and escape across the sample. The ratio between the intensity of the incident and escaped rays defines the absorbance of light by the sample. This value of light absorbance is of main interest in UV-VIS spectroscopy. As in Equation (2), light absorbance is dependent on ε, c, and l [18]. The absorbance value(s) at a single or multiple wavelength(s) will then be used to measure the concentration of compounds in a sample [19][20][21][22][23].
Molecules 2019, 24, x FOR PEER REVIEW 4 of 39 Equation 2, light absorbance is dependent on ε, c, and l [18]. The absorbance value(s) at a single or multiple wavelength(s) will then be used to measure the concentration of compounds in a sample [19][20][21][22][23].
Fluorescence Spectroscopy
Fluorescence spectroscopy is distinct from other spectroscopy techniques in terms of the emission of light when incident rays from an ultraviolet or visible light source is absorbed by fluorescent molecules present in a sample. These fluorescent molecules are known as fluorophores and commonly known examples include quinine, fluorescein, acridine orange, rhodamine B and pyridine 1 [24].
The fluorescence phenomenon can be explained with a Jablonski diagram illustrated in Figure 2. It should first be understood that fluorescence involves the three electronic states of a fluorophore molecule, namely the singlet ground, first and second electronic states. These states are represented by S0, S1 and S2 in Figure 2. The key condition for fluorescence to occur is the excitation of the molecule from the ground state, S0 to either electronic states S1 or S2 upon the absorption of light. If the molecule reaches the S2 state, internal conversion or vibrational relaxation will occur, returning the molecule to the lower S1 state without radiation emitted. From here, the molecule will again return to the S0 while emitting light which has equal energy as the energy difference between S0 and S1. This light emission is known as fluorescence and this condition typically occurs 10 -8 seconds after the initial excitation [17]. Fluorescence spectroscopy is highly specific and highly sensitive. The high specificity of the technique arises from the usage of both the excitation and emission spectra; whereas high sensitivity is achieved as radiation measurements are made against absolute darkness. These characteristics however limit the independent usage of the technique [17]. As a result, fluorescence spectroscopy is often combined with high performance liquid chromatography (HPLC) [25]. Variations may also be implemented in the excitation and emission wavelengths, forming the synchronous fluorescence spectroscopy (SFS) [26].
Fluorescence Spectroscopy
Fluorescence spectroscopy is distinct from other spectroscopy techniques in terms of the emission of light when incident rays from an ultraviolet or visible light source is absorbed by fluorescent molecules present in a sample. These fluorescent molecules are known as fluorophores and commonly known examples include quinine, fluorescein, acridine orange, rhodamine B and pyridine 1 [24].
The fluorescence phenomenon can be explained with a Jablonski diagram illustrated in Figure 2. It should first be understood that fluorescence involves the three electronic states of a fluorophore molecule, namely the singlet ground, first and second electronic states. These states are represented by S 0 , S 1 and S 2 in Figure 2. The key condition for fluorescence to occur is the excitation of the molecule from the ground state, S 0 to either electronic states S 1 or S 2 upon the absorption of light. If the molecule reaches the S 2 state, internal conversion or vibrational relaxation will occur, returning the molecule to the lower S 1 state without radiation emitted. From here, the molecule will again return to the S 0 while emitting light which has equal energy as the energy difference between S 0 and S 1 . This light emission is known as fluorescence and this condition typically occurs 10 -8 seconds after the initial excitation [17]. , light absorbance is dependent on ε, c, and l [18]. The absorbance value(s) at a single or multiple wavelength(s) will then be used to measure the concentration of compounds in a sample [19][20][21][22][23].
Fluorescence Spectroscopy
Fluorescence spectroscopy is distinct from other spectroscopy techniques in terms of the emission of light when incident rays from an ultraviolet or visible light source is absorbed by fluorescent molecules present in a sample. These fluorescent molecules are known as fluorophores and commonly known examples include quinine, fluorescein, acridine orange, rhodamine B and pyridine 1 [24].
The fluorescence phenomenon can be explained with a Jablonski diagram illustrated in Figure 2. It should first be understood that fluorescence involves the three electronic states of a fluorophore molecule, namely the singlet ground, first and second electronic states. These states are represented by S0, S1 and S2 in Figure 2. The key condition for fluorescence to occur is the excitation of the molecule from the ground state, S0 to either electronic states S1 or S2 upon the absorption of light. If the molecule reaches the S2 state, internal conversion or vibrational relaxation will occur, returning the molecule to the lower S1 state without radiation emitted. From here, the molecule will again return to the S0 while emitting light which has equal energy as the energy difference between S0 and S1. This light emission is known as fluorescence and this condition typically occurs 10 -8 seconds after the initial excitation [17]. Fluorescence spectroscopy is highly specific and highly sensitive. The high specificity of the technique arises from the usage of both the excitation and emission spectra; whereas high sensitivity is achieved as radiation measurements are made against absolute darkness. These characteristics however limit the independent usage of the technique [17]. As a result, fluorescence spectroscopy is often combined with high performance liquid chromatography (HPLC) [25]. Variations may also be implemented in the excitation and emission wavelengths, forming the synchronous fluorescence spectroscopy (SFS) [26]. Fluorescence spectroscopy is highly specific and highly sensitive. The high specificity of the technique arises from the usage of both the excitation and emission spectra; whereas high sensitivity is achieved as radiation measurements are made against absolute darkness. These characteristics however limit the independent usage of the technique [17]. As a result, fluorescence spectroscopy is often combined with high performance liquid chromatography (HPLC) [25]. Variations may also be implemented in the excitation and emission wavelengths, forming the synchronous fluorescence spectroscopy (SFS) [26].
Infrared (IR) Spectroscopy
Infrared (IR) spectroscopy operates within the IR band with wavelengths from 780 nm to 1 mm. The IR band can be further broken down into three sub-bands, namely near-infrared (NIR; 780 nm to 5 µm), mid-infrared (MIR; 5 µm to 30 µm) and far-infrared (FIR; 30 µm to 1 mm). In agriculture-related optics and photonics, the NIR and MIR bands are of greater interest [17].
IR spectroscopy obtains the spectral information of a subject due to molecular vibrations under the excitation of an IR light source. In general, molecular vibrations occur when there exist normal modes of vibrations. A normal mode of vibration (or fundamental) refers to the phenomenon in which every atom in a molecule experiences a simple harmonic oscillation about its equilibrium position. These atoms oscillate in phase at the same frequency while the center of gravity of the molecule remains unchanged. A typical molecule has 3N-6 fundamentals (3N-5 for linear molecules), where N refers to the number of atoms. The diatomic molecular vibrations are illustrated in Figure 3 [27]. Infrared (IR) spectroscopy operates within the IR band with wavelengths from 780 nm to 1 mm. The IR band can be further broken down into three sub-bands, namely near-infrared (NIR; 780 nm to 5 µm), mid-infrared (MIR; 5 µm to 30 µm) and far-infrared (FIR; 30 µm to 1 mm). In agriculturerelated optics and photonics, the NIR and MIR bands are of greater interest [17].
IR spectroscopy obtains the spectral information of a subject due to molecular vibrations under the excitation of an IR light source. In general, molecular vibrations occur when there exist normal modes of vibrations. A normal mode of vibration (or fundamental) refers to the phenomenon in which every atom in a molecule experiences a simple harmonic oscillation about its equilibrium position. These atoms oscillate in phase at the same frequency while the center of gravity of the molecule remains unchanged. A typical molecule has 3N-6 fundamentals (3N-5 for linear molecules), where N refers to the number of atoms. The diatomic molecular vibrations are illustrated in Figure 3 Molecular vibrations, which occur regardless the presence of IR light source, result in an increase in light absorption. These peaks in absorption form specific bands in the IR spectrum that correspond to the specific frequencies in which molecular vibrations occur. This allows the easy identification of the molecular structure in a sample since different molecules have different vibration frequencies [27]. This unique frequency 'fingerprint' is exceptionally beneficial in the analysis of complex molecules that contains functional groups such as -OH, -NH2, -CH3, C=O, C6H5-and more. For instance, the C6H5-group forms peaks at wavenumbers from 1600 cm −1 to 1500 cm −1 (wavelengths from 6.25 µm to 6.67 µm) whereas the C=O group exhibits high absorption at wavenumbers from 1800 cm −1 to 1650 cm −1 (wavelengths from 5.56 µm to 6.06 µm) [28].
Near-infrared (NIR) Spectroscopy
The near-infrared (NIR) spectroscopy operates within the NIR band with wavelengths from 780 nm to 5 µm. The absorptions within the NIR band exist due to overtones and combinations of the fundamental vibrations. Overtones refer to higher frequencies that are multiples of the fundamental frequency. Meanwhile, combinations involve interactions between two or more vibrations occurring simultaneously, resulting in a frequency which is the sum of multiples of the respective frequencies.
A majority of the absorptions in the NIR band are due to vibrations of the C-H, O-H and N-H bands. The S-H and C=O bonds too potentially contribute to these absorptions. Several assignments of the NIR absorption bands can be seen in Table 1 [29]. Table 1. Examples of NIR absorption bands [29]. Molecular vibrations, which occur regardless the presence of IR light source, result in an increase in light absorption. These peaks in absorption form specific bands in the IR spectrum that correspond to the specific frequencies in which molecular vibrations occur. This allows the easy identification of the molecular structure in a sample since different molecules have different vibration frequencies [27]. This unique frequency 'fingerprint' is exceptionally beneficial in the analysis of complex molecules that contains functional groups such as -OH, -NH 2 , -CH 3 , C=O, C 6 H 5 -and more. For instance, the C 6 H 5 -group forms peaks at wavenumbers from 1600 cm −1 to 1500 cm −1 (wavelengths from 6.25 µm to 6.67 µm) whereas the C=O group exhibits high absorption at wavenumbers from 1800 cm −1 to 1650 cm −1 (wavelengths from 5.56 µm to 6.06 µm) [28].
Near-Infrared (NIR) Spectroscopy
The near-infrared (NIR) spectroscopy operates within the NIR band with wavelengths from 780 nm to 5 µm. The absorptions within the NIR band exist due to overtones and combinations of the fundamental vibrations. Overtones refer to higher frequencies that are multiples of the fundamental frequency. Meanwhile, combinations involve interactions between two or more vibrations occurring simultaneously, resulting in a frequency which is the sum of multiples of the respective frequencies.
A majority of the absorptions in the NIR band are due to vibrations of the C-H, O-H and N-H bands. The S-H and C=O bonds too potentially contribute to these absorptions. Several assignments of the NIR absorption bands can be seen in Table 1 [29]. NIR spectroscopy, which is a non-destructive measurement, enables the simultaneous identification of components in a single sample within a short period of time, making it a preferable replacement for various chemical techniques. However, consideration should be taken into account as this technique requires initial calibration with samples of known composition, requiring great expenses of time and resources. Not least, frequent recalibration and issue of instrument interoperability might affect the practicality of the NIR spectroscopy technique [29].
Mid-Infrared (MIR) Spectroscopy
The mid-infrared (MIR) spectroscopy operates within the MIR band with wavelengths from 5 µm to 30 µm (wavenumbers from 4000 cm −1 to 400 cm −1 ; note the presence of slight overlapping with NIR). The absorptions that occur within the MIR band are due to fundamental vibrations and can be segregated into four regions, namely the X-H stretching region (4000 cm −1 to 2500 cm −1 ), triple-bond region (2500 cm −1 to 2000 cm −1 ), double-bond region (2000 cm −1 to 1500 cm −1 ) as well as the fingerprint region (1500 cm −1 to 600 cm −1 ) [27].
The X-H stretching region is due to vibrations from O-H, C-H and N-H stretching. The triple-bond region arises from vibrations of C≡C and C≡N bonds. Besides, the double-bond region relates to C=C, MIR spectroscopy is effective since it provides information on structure-function relationships while performing quantitative analysis. The structure-function relationships are useful in food research and quality control, making MIR spectroscopy a crucial technique in the field of agriculture. The Fourier transform process is often bundled with MIR spectroscopy for data analysis, forming the popular Fourier transform infrared spectroscopy (FTIR) technique [27].
Raman Spectroscopy
Raman spectroscopy (RS), similar to IR spectroscopy, is another form of vibrational spectroscopy technique. RS obtains the spectral information of samples due to the occurrence of Raman effects [30]. Prior to understanding the Raman effects, one should look into the light scattering schemes that occur when incident photons interact with molecules in the sample. The possible light scattering schemes are illustrated in Figure 4. In the case of elastic scattering or Rayleigh scattering, the excited photons experience no change in energy content upon returning to ground state. Alternately, in the case of inelastic scattering or Raman scattering, the excited photons may lose (Stokes' shift) or gain (Anti-Stokes' shift) energy equivalent to the vibrational energy changes in the atoms of the molecules. This affects the motion of the atoms as well as the polarizability of the molecule. The change in molecule polarizability results in increased Raman intensity, ultimately forming the Raman spectrum when plotted across the investigated wavenumbers. However, this effect is weak as the probability of energy exchange is low [30].
The RS technique is gaining popularity as it enables the identification of molecular structure through the characteristic wavenumber in which vibrations occur. Furthermore, samples can be studied in the absence of a solvent as water causes weak Raman scattering. Not least, this technique is instantaneous and may undergo intensity enhancement. However, this technique is not without limitations. Due to the low probability of Raman scattering, this technique requires high concentration of samples. Moreover, sample molecules may experience photo degradation due to excitation of electronic absorption bands. The existence of fluorescence from impurities may disrupt the results obtained as well. These limitations aside, the RS technique can be combined with IR spectroscopy to deliver satisfactory results as summarized in Table 3 [30].
when incident photons interact with molecules in the sample. The possible light scattering schemes are illustrated in Figure 4. In the case of elastic scattering or Rayleigh scattering, the excited photons experience no change in energy content upon returning to ground state. Alternately, in the case of inelastic scattering or Raman scattering, the excited photons may lose (Stokes' shift) or gain (Anti-Stokes' shift) energy equivalent to the vibrational energy changes in the atoms of the molecules. This affects the motion of the atoms as well as the polarizability of the molecule. The change in molecule polarizability results in increased Raman intensity, ultimately forming the Raman spectrum when plotted across the investigated wavenumbers. However, this effect is weak as the probability of energy exchange is low [30]. Apart from the popular spectroscopy techniques discussed earlier, existing studies presented additional variations of spectroscopy techniques which may be more complex in nature. For instance, dielectric spectroscopy has been utilized in agricultural inspections. Dielectric spectroscopy involves the inspection of dielectric properties or permittivity of samples over broad frequency ranges. Dielectric properties or permittivity refers to the ability of samples to store electrical energy in the electric field. In this spectroscopy technique, the permittivity is a complex permittivity relative to the free space, and this complex number is represented by (3): where the real part, ε , is the dielectric constant and the imaginary part, ε", is the dielectric loss factor which covers losses due to dipolar relaxation and ionic conduction [31]. Another spectroscopy variation is the nuclear magnetic resonance (NMR) spectroscopy technique. The NMR spectroscopy gains spectral information of samples from the interaction between the magnetic moments of nuclei of various atoms and the applied magnetic fields. The two common phenomena that give rise to the NMR spectra are chemical shift and J-coupling [32]. A chemical shift occurs due to different resonant frequencies present in nuclei of the same species. The difference in resonant frequencies is a result of shielding effect from electrons surrounding the nuclei. The shielding effect is sensitive to chemical environments, hence allowing the characteristic identification of specific molecular functional groups [32].
The J-coupling phenomenon is also known as indirect (scalar) spin-spin coupling. This coupling effect results in splitting of spectroscopic lines into multiplets. The J-coupling occurs between two nuclei or groups of nuclei and is governed by the polarization of electrons on the chemical bonds connecting these nuclei. The polarization scheme is in turn dependent on the instant orientation of the nuclear magnetic moments in the presence of a magnetic field [32].
Spectroscopy Processing and Analysis
The raw spectral data undergoes pre-processing or pre-treatment in order to reduce noise and correct baseline variations. The common pre-treatment techniques are multiplicative scattering correction (MSC), standard normal variate (SNV), Savitzky-Golay smoothing as well as first and second derivatives [33,34].
Upon the completion of pre-processing or pre-treatment, the data set undergoes multivariate analysis to select and extract wavelengths that contain useful information. This aids in rectifying issues of collinearity, band overlapping and interaction between spectral variables. The results from multivariate analysis will be used to develop calibration models for calibration and prediction purposes [33,34].
The developed calibration models can be categorized according to the nature of the utilized multivariate analysis such as linear regression or nonlinear regression. Calibration models based on linear regression are built from partial least squares (PLS), interval partial least squares (iPLS), synergy interval partial least squares (SiPLS) or successive projections algorithm (SPA). Meanwhile, calibration models based on nonlinear regression are constructed from principal component analysis (PCA), independent component analysis (ICA), support vector machines (SVM), artificial neural networks (ANN) or a genetic algorithm (GA) [33,34].
The robustness of the final calibration model is evaluated from its ability to perform calibration and prediction. The calibration performance of the model is determined from the root mean square error of calibration (RMSEC) and the correlation coefficient (R C ) in the calibration set. Meanwhile, the prediction performance of the model is identified from the root mean square error of prediction (RMSEP) and the correlation coefficient (R P ) in the prediction set. Ideally, an effective model should register low RMSEC and RMSEP, with minimum difference between RMSEC and RMSEP. Not least, higher R C and R P are preferable [33,34].
Spectral Imaging Technique
The spectral imaging technique is a combination of both imaging and spectroscopy techniques discussed earlier. Being a combinational technique, the spectral imaging technique preserves the best of both worlds, allowing the simultaneous extraction of spatial and spectral information from the inspected sample [35,36].
Classes of Spectral Imaging
The spectral imaging technique acquires multiple images of the same subject at varying wavelengths. The resulting spectral images are three-dimensional (3-D) in nature, consisting of two spatial dimensions (row, x, and column, y) and one spectral dimension (wavelength, λ). Variations of spectral imaging technique are determined by the continuity of data in the wavelength dimension, branching out into hyperspectral imaging and multispectral imaging [35,36].
In general, hyperspectral imaging obtains spectral images in continuous wavelengths, whereas multispectral imaging registers spectral images at discrete wavelengths. Hyperspectral imaging acquires large number of images at high spatial and spectral resolutions. Due to the high volume of data, hyperspectral imaging requires long image acquisition time and involves complex algorithms for image analysis. Despite the complexity, hyperspectral imaging is essential for fundamental research and is the basis for multispectral imaging [35,36].
Multispectral imaging acquires spectral images at a significantly smaller number compared to hyperspectral imaging. Spectral images will only be acquired at optimal wavelengths predetermined from the analysis of dataset obtained through hyperspectral imaging. A smaller number of interested wavelengths allows rapid image acquisition and requires simpler image analysis algorithms. This characteristic of optimum data volume makes multispectral imaging perfectly suited for real-time in-field applications [35,36].
Spectral Image Acquisition Methods
There are several methods in which spectral imaging systems acquire spectral images. The methods are point scan, line scan and area scan as illustrated in Figure 5 [35]. The point scan (whiskbroom) method acquires the spectrum of a single pixel in each scan. A complete hyperspectral cube will be generated as the detector moves from pixel to pixel along the two spatial axes (x and y). The point scan method is similar to a normal spectroscopic approach. Since it cannot cover a large sample area, the point scan method is time consuming and unsuitable for fast image acquisition [35,36]. acquires large number of images at high spatial and spectral resolutions. Due to the high volume of data, hyperspectral imaging requires long image acquisition time and involves complex algorithms for image analysis. Despite the complexity, hyperspectral imaging is essential for fundamental research and is the basis for multispectral imaging [35,36]. Multispectral imaging acquires spectral images at a significantly smaller number compared to hyperspectral imaging. Spectral images will only be acquired at optimal wavelengths predetermined from the analysis of dataset obtained through hyperspectral imaging. A smaller number of interested wavelengths allows rapid image acquisition and requires simpler image analysis algorithms. This characteristic of optimum data volume makes multispectral imaging perfectly suited for real-time infield applications [35,36].
Spectral Image Acquisition Methods
There are several methods in which spectral imaging systems acquire spectral images. The methods are point scan, line scan and area scan as illustrated in Figure 5 [35]. The point scan (whiskbroom) method acquires the spectrum of a single pixel in each scan. A complete hyperspectral cube will be generated as the detector moves from pixel to pixel along the two spatial axes (x and y). The point scan method is similar to a normal spectroscopic approach. Since it cannot cover a large sample area, the point scan method is time consuming and unsuitable for fast image acquisition [35,36]. The line scan (pushbroom) method, in each scan, acquires a slit (line) of spatial information together with the spectrum of every pixel along the line. A complete hyperspectral cube will be formed when scans are repeated along the direction of motion (x). The operation characteristic of the line scan method makes it suitable to acquire spectral images of moving samples. Hence, this method is usually combined with conveyor belt systems, making it a popular method in practical production lines. However, the exposure time should be short and accurately selected to allow uniform exposure at all wavelengths [35,36].
The area scan (band sequential) method, on the other hand, acquires a 2-D grayscale image comprising of complete spatial information in a single wavelength. A complete hyperspectral cube is generated through image stacking when scans are performed along the spectral axis (λ). The nature of the area scan method makes it more suited for the imaging of stationary samples instead of moving samples. In short, among the image acquisition methods discussed, line scan and area scan are greatly preferred over point scan for both hyperspectral and multispectral imaging on the basis of time The line scan (pushbroom) method, in each scan, acquires a slit (line) of spatial information together with the spectrum of every pixel along the line. A complete hyperspectral cube will be formed when scans are repeated along the direction of motion (x). The operation characteristic of the line scan method makes it suitable to acquire spectral images of moving samples. Hence, this method is usually combined with conveyor belt systems, making it a popular method in practical production lines. However, the exposure time should be short and accurately selected to allow uniform exposure at all wavelengths [35,36].
The area scan (band sequential) method, on the other hand, acquires a 2-D grayscale image comprising of complete spatial information in a single wavelength. A complete hyperspectral cube is generated through image stacking when scans are performed along the spectral axis (λ). The nature of the area scan method makes it more suited for the imaging of stationary samples instead of moving samples. In short, among the image acquisition methods discussed, line scan and area scan are greatly preferred over point scan for both hyperspectral and multispectral imaging on the basis of time consumption [35,36].
Spectral Imaging Sensing Modes
Spectral imaging may have varying sensing modes as illustrated in Figure 6. The sensing modes are determined by the positions of the light source and the detector, forming variations such as reflectance, transmittance and interactance modes. In reflectance mode, the detector collects the light reflected off the illuminated surface. This sensing mode is suitable for identifying external features of samples such as size, shape, color, texture and defects. However, when selecting this mode, the detector should be properly positioned to avoid specular reflection [36]. of samples such as size, shape, color, texture and defects. However, when selecting this mode, the detector should be properly positioned to avoid specular reflection [36]. The transmittance mode operates by having the detector collect light rays transmitted through inspected samples. In this sensing mode, the light source and the detector will be placed in opposite direction to each other. Due to the absorption of light rays in a sample, the detected signal will be relatively weak and dependent on sample thickness. Hence, the transmittance mode is commonly applied in the internal inspection of relatively transparent samples [36].
Meanwhile, the interactance mode overcomes the limitations of both the reflectance and transmittance modes. This sensing mode exhibits less surface effect compared to reflectance mode. At the same time, it allows detection in deeper layers of a sample without being affected by sample thickness as in transmittance mode. This advantageous sensing mode is set up by installing the light source and the detector at the same side and parallel to each other [36].
Spectral Imaging System Construction
The variations in spectral imaging lead to a diversity of instruments during the construction of a spectral imaging system. In general, a spectral imaging system is made up of a light source, a wavelength dispersive device and an area detector [35,36].
The light source for a spectral imaging system can be classified into illumination and excitation sources. Illumination light source is selected when measurements involve changes in the intensity of the incident rays upon light-sample interaction. The spectral composition of the incident source will not experience any changes. Such interaction is commonly observed in reflectance and transmittance sensing modes. Broadband lights are normally used as illumination sources. An example of illumination light source is the quartz tungsten halogen (QTH) lamp which is capable of generating a smooth spectrum in the visible to infrared range. Besides, the broadband light emitting diode (LED) has gained popularity over time due to its low power consumption, low heat generation, small size The transmittance mode operates by having the detector collect light rays transmitted through inspected samples. In this sensing mode, the light source and the detector will be placed in opposite direction to each other. Due to the absorption of light rays in a sample, the detected signal will be relatively weak and dependent on sample thickness. Hence, the transmittance mode is commonly applied in the internal inspection of relatively transparent samples [36].
Meanwhile, the interactance mode overcomes the limitations of both the reflectance and transmittance modes. This sensing mode exhibits less surface effect compared to reflectance mode. At the same time, it allows detection in deeper layers of a sample without being affected by sample thickness as in transmittance mode. This advantageous sensing mode is set up by installing the light source and the detector at the same side and parallel to each other [36].
Spectral Imaging System Construction
The variations in spectral imaging lead to a diversity of instruments during the construction of a spectral imaging system. In general, a spectral imaging system is made up of a light source, a wavelength dispersive device and an area detector [35,36].
The light source for a spectral imaging system can be classified into illumination and excitation sources. Illumination light source is selected when measurements involve changes in the intensity of the incident rays upon light-sample interaction. The spectral composition of the incident source will not experience any changes. Such interaction is commonly observed in reflectance and transmittance sensing modes. Broadband lights are normally used as illumination sources. An example of illumination light source is the quartz tungsten halogen (QTH) lamp which is capable of generating a smooth spectrum in the visible to infrared range. Besides, the broadband light emitting diode (LED) has gained popularity over time due to its low power consumption, low heat generation, small size and long lifetime [35,36].
Excitation light source is usually selected when measurements involve changes in the frequency and wavelength of the incident rays. Interactions of this nature usually involve fluorescence phenomenon or Raman scattering effect. Narrowband lights are frequently used as excitation sources. A popular excitation light source is the laser which generates powerful monochromatic rays. Not least, UV fluorescent lamp, narrowband LED, high-pressure arc lamp (xeon arc lamp) and low-pressure vapor lamp (mercury vapor lamp) add to the family of excitation light sources [30,31].
The core component of the spectral imaging system is the wavelength dispersive device. A wavelength dispersive device disperses broadband light into different wavelengths to be projected to the area detector. Examples of wavelength dispersive devices include the imaging spectrograph, electronically tunable filter and beam splitting device [35,36].
Compared to traditional spectrograph, an imaging spectrograph extracts both spatial and spectral information. The imaging spectrograph disperses the broadband light illuminated onto different spatial areas of a sample into different wavelengths. This is achieved through diffraction gratings. The two most popular imaging spectrographs are the prism-grating-prism (PGP) imaging spectrograph which uses transmission diffraction gratings and the Offner imaging spectrograph that uses reflection diffraction gratings [35,36]. These variations of imaging spectrographs are commonly applied in line scan acquisitions [37].
An electronically tunable filter utilizes electronic devices to extract the required wavelength. Current electronically tunable filters can be categorized into the acousto-optic tunable filter (AOTF) and liquid crystal tunable filter (LCTF). An AOTF utilizes an acoustic transducer to generate high frequency acoustic waves that change the refractive index of a crystal. The crystal with varied refractive index will only allow the passage of light rays at the specified wavelength. Meanwhile, a LCTF transmits light at the required wavelength through electronically controlled liquid crystal cells [35,36]. These electronically tunable filters allow fast and flexible wavelength switching compared to mechanical filter wheels. They too exhibit advantages of high optical throughput, narrow bandwidth and broad spectral range [38].
Unlike the electronically tunable filter, a beam splitting device allows spectral images to be obtained simultaneously at multiple wavelengths. The beam splitting device divides light into several parts and passes them through bandpass filters which correspond to the required wavelengths. The beam splitting device can be categorized into color splitting and neutral splitting. In color splitting, light rays at a particular waveband are directed to each output, whereas, in neutral splitting, an equal portion of the total light energy is directed to each output [35]. The multiple wavelength acquisition characteristic makes the beam splitting device suitable to be installed in multispectral imaging systems [39]. A spectral imaging system will not be complete without an area detector. The area detector is responsible for collecting light rays which will eventually form the spectral images of the inspected sample. The common categories of area detector are the charge-couple device (CCD) camera and the complementary metal-oxide-semiconductor (CMOS) camera [35,36].
A CCD camera is made up of millions of photodiodes (pixels) that are closely arranged to form an array. These light sensitive photodiodes convert the incident photons into electric charges that correspond to the intensity of the exposed incident rays. The accumulated electric charges at each photodiode will then be moved out of the array to be quantified for spectral image formation [35,36]. One of the common CCD cameras is the silicon CCD camera. The silicon CCD camera exploits the sensitivity of silicon under visible light to perform image acquisition in visible and short-wavelength near-infrared bands [40]. Indium gallium arsenide (InGaAs) CCD camera is another CCD camera variation constructed from InGaAs, an alloy between indium arsenide (InAs) and gallium arsenide (GaAs) which is sensitive in the near-infrared band [41]. Not least, mercury cadmium telluride (MCT or HgCdTe) CCD camera built from HgCdTe, an alloy between mercury telluride (HgTe) and cadmium telluride (CdTe), enables sensing in the long-wavelength near-infrared and mid-infrared band [42].
Comparatively, the CMOS camera is similar to the CCD camera by having a collection of photodiodes (pixels) that convert light rays into electrical charges. The difference, however, lies in the quantification process of the electric charges. Opposed to the remote quantification in the CCD camera, the CMOS camera allows electric charges at each pixel to be independently and instantaneously read by the transistor attached to each photodiode [43]. This unique characteristic allows the CMOS camera to compete with the CCD camera in terms of high imaging acquisition speed, blooming immunity, low cost, low power consumption and small size. However, careful note should be taken as the CMOS camera is susceptible to noise due to on-chip signal transmissions, resulting in lower sensitivity and dynamic range when pitted against CCD camera [36].
Spectral Imaging Processing and Analysis
The raw spectral image data obtained via the spectral imaging technique comes in different formats according to the image acquisition method used. The common formats are Band interleaved by pixel (BIP), band interleaved by line (BIL) and band sequential (BSQ). The BIP format results from the point scan method and stores the complete spectrum of each pixel sequentially. The BIL format comes with the line scan method and stores the complete spectrum of each line in order. Lastly, the BSQ format relates to the area scan method and stacks the spatial image continuously obtained at each wavelength [35,36].
Similar to the imaging and spectroscopy techniques, the raw spectral image data in BIP, BIL and BSQ formats should undergo pre-processing in both the spatial and spectral aspects before being utilized for further analysis. The raw spectral image, which represents detector signal intensity, will first undergo flat-field calibration or reflectance calibration to form useful reflectance or absorbance image. From the spatial aspect, the generated reflectance image can be further improved through image enhancement processes such as edge and contrast enhancement, magnifying, pseudo-coloring and sharpening. Noise reduction can also be achieved through spatial filtering, Fourier transform (FT) and wavelet transform (WT). From the spectral aspect, noise reduction and baseline correction can be performed through algorithms such as MSC, SNV, Savitzky-Golay smoothing, first and second derivatives, FT, WT as well as orthogonal signal correction (OSC) [35,36]. The next step in the analysis flow will be image segmentation. Image segmentation serves to divide the pre-processed spectral image into different regions for the identification of region of interests (ROIs) [44]. In this process, segmentation algorithms are greatly preferred over manual segmentation due to the ease of operation and time saving. The selections of segmentation algorithms include thresholding (global thresholding or adaptive thresholding), morphological processing (erosion, dilation, open, close or watershed algorithm), edge-based segmentation (gradient-based or Laplacian-based methods) and spectral image segmentation [36].
Spatial analysis utilizing spectral image data usually involves quantitative measurement. In this process, gray-level object measurement is performed to quantify the intensity distribution of ROI extracted from image segmentation. Gray-level object measurements can be categorized according to intensity-based or texture-based measurements [45]. Intensity-based measurements are usually first-order measures such as mean [46,47], standard deviation, skew, energy and entropy [36]. Meanwhile, texture-based measurements are second-order measures such as joint distribution functions [36], gray-level co-occurrence matrix (GLCM) [46,48] and 2-D Gabor filter [49].
For spectral analysis, the data set will undergo multivariate analysis to reduce the spectral dimension and select the optimum wavelengths. Similar to the spectroscopy technique, some examples of multivariate analysis algorithms include PLS, linear discrimination analysis (LDA) [35,36], correlation analysis (CA) [50], PCA, ICA [41,51,52], ANN [53], sequential forward selection (SFS) [54] and GA [55]. These results from multivariate analysis will be used to develop calibration models for calibration, validation and prediction purposes [36].
The robustness of the final calibration model is evaluated from its ability to perform calibration and prediction. The calibration performance of the model is determined from the standard error of calibration (SEC), root mean square error of calibration (RMSEC) and the coefficient of determination (r 2 C ) in the calibration set. Validation performance is determined via the root mean square error of cross-validation (RMSECV) and the coefficient of determination (r 2 V ) in the validation set. Meanwhile, the prediction performance of the model is identified from standard error of prediction (SEP), root mean square error of prediction (RMSEP), residual predictive deviation (RPD) and the coefficient of determination (r 2 P ) in the prediction set. Ideally, an effective model should register low SEC, RMSEC, RMSECV, SEP and RMSEP, with a minimum difference between SEC and SEP. Not least, higher r 2 C , r 2 P , r 2 P and RPD are preferable [36].
Pros and Cons of Spectral Imaging
This technique is advantageous as it omits chemical processes and requires minimum sample preparation. Moreover, the composition of multiple components in a sample can be simultaneously obtained. Upon spectral image acquisition, spectral imaging too allows the flexible selection of region of interest (ROI) for analysis. Furthermore, owing to the rich spatial and spectral information, spectral imaging can easily detect and differentiate subjects even though similar colors, overlapping spectra and morphological characteristics are present [36].
However, the spectral imaging technique does pose several limitations. Hardware speed is a major concern, especially in the case of hyperspectral imaging, due to the massive amounts of data to be acquired and analyzed. Moreover, spectral imaging includes the acquisition of redundant data, resulting in complex data analysis. Spectral imaging systems too require constant calibration in order to maintain their efficiency. The detection limits of spectral imaging are poorer compared to chemical-based analytical methods. Similar to spectroscopy, spectral imaging suffers from multicollinearity and requires multivariate analysis to address the issue. In addition, spectral imaging is inapplicable when the ROI is smaller than the size of a pixel or does not exhibit the characteristic spectral absorption. Lastly, spectral imaging may be irrelevant in the analysis of liquids and homogeneous samples since these samples do not pose distinctive and useful spatial information [36]. Table 4 presents a simple comparison of the optics and photonics techniques in agriculture that have been discussed earlier. From the comparison, the imaging technique is noted to be utilized for the extraction of spatial information only and is sensitive to small-sized objects. In contrast to the imaging technique, the spectroscopy technique allows acquisition of spectral information and is useful in accessing multi-constituent information. The spectral imaging technique covers the benefits of both imaging and spectroscopy techniques, allowing it to obtain spectral and spatial information simultaneously. Apart from this, spectral imaging has the added value of flexible spectral extraction as well as the capability of generating quality-attribute distribution. However, it should be noted that multispectral imaging has poorer access to spectral information compared to hyperspectral imaging due to the acquisition at limited number of wavelengths. Among the compared techniques, the spectral imaging technique can be said to be the most robust. Nonetheless, the area of application should be given the utmost consideration when selecting the best optics and photonic technique in order to avoid the underutilization or overutilization of a particular technique [36]. Table 4. Comparison of optics and photonics techniques in agriculture [36].
Characteristics
Imaging Spectroscopy Spectral Imaging
Multi-constituent information ×
Sensitivity to small-sized objects ×
Flexibility of spectral extraction × ×
Generation of quality-attribute distribution × ×
Optics and Photonics Applications in Agriculture
The optics and photonics techniques discussed above have been applied in various studies involving agricultural products. The studies will be tabulated in the following sections, enlisting details such as agriculture class, agriculture product, application area, wavelength details and country of applications. Table 5 lists some of the agricultural works based on the imaging technique. The imaging technique is performed in the UV-VIS-IR range and involves the acquisition of spatial, color and thermal data from the inspected samples. These works show that the imaging technique is suited for inspection or analysis based on external features of the subject of interest. For instance, bruise detection [56,57] and disease detection [58,59] are performed by inspecting the external damage on the sample. In addition, quantitative analysis [60,61] is performed using the spatial information obtained. The color features extracted are also used for maturity evaluation [57,62,63] and nutrient content detection [64,65]. The thermal data, meanwhile, proves to be useful in similar occasions of bruise detection [66,67], disease detection [68,69] and maturity evaluation [70,71] by analyzing the temperature variations over the inspected sample. Not least, the most significant application of the imaging technique is the development of automated agricultural robots [72][73][74][75] and animal behavioral studies [76,77]. Apple Yield estimation (thermal) [79] Apple Scab disease detection (thermal) [68] Green apple Acquisition of segmented fruit region [80] Green apple and orange Yield estimation [61] Orange Texture analysis [81] Orange Bruise detection (thermal) [67] Citrus Water stress evaluation (thermal) [82] Pear Maturity evaluation (thermal) [71] Banana Maturity evaluation [62] Banana Maturity evaluation [63] Persimmon Maturity evaluation (thermal) [71] Passion fruit Mass and volume estimation [83] Blueberry Bruise detection [56] Grapevine Pathogen detection (thermal) [84] Tomato Fruit detection [85,86]
Tomato
Bruise detection and maturity evaluation [57] Tomato Bruise detection (thermal) [87] Tomato Maturity evaluation (thermal) [71] Tomato Clustered fruit detection [88] Sweet peppers Peduncle detection [89] Onion Post-harvest quality assessment (thermal) [ Farm and Plantation Seed Viability evaluation (thermal) [104] Wheat field Estimation of nutrient content [65] Cauliflower plantation Weed detection [73] Asparagus plantation Crop harvest robot vision [74] Sugar beet and rape plantation Agriculture robot vision [75] Grapevines Estimation of intra-parcel grape quantities [105] Cow farm Behavioural studies [76,106] Goat and sheep farm Animal species identification [107] Fish aquarium Behavioural studies [77,108] Baby shrimp farm Chlorine level detection [109] Orchid farm Disease and pest detection [58] Surface and ground water Chemical content detection [110] Based on Table 5, bruise detection, yield estimation, and disease identification are the three most common applications with imaging technique in agriculture. In bruise detection, a hyperspectral camera with broad operating wavelength from 400 to 5000 nm [78], a non-destructive and non-contact infrared sensing thermogram [66], and an infrared thermal imaging camera with high temperature resolution of 0.1 K [70] are among the instruments employed as the imaging technique. Moreover, the thermal camera with temperature resolution better than 0.5 • C [79], colour stereo vision camera which creates a 3D environment for further processing [86], and grading machine with a high accuracy of 96.47% [57] are employed for yield estimation. However, the grading machine proposed in [57] has a small capacity in estimation for 300 tomatoes per hour and does not efficiently work for tomato images with high specular reflection. In addition, infrared thermography is a popular device for disease identification due to its non-invasive monitoring and indirect visualization of downy mildew development [69]. This device takes the colour reflectance image for the detection of V. inaequalis development on apple leaves [68] and detects the pathogen in grapevines [84]. Additionally, an X-ray computed tomography scanner is utilized to obtain the cross-section of onion inoculated by pathogens [90], whilst an unmanned aerial vehicle (UAV) is presented to track the foliar disease in soybean [98].
Apart from the instrument, numerous types of algorithms are depicted in imaging technique. In bruise detection, PCA and a minimum noise fraction are proposed for 20 apple samples with threshold percentages of success within 86% to 93% and 87% to 97%, respectively [78]. In yield estimation, the fruit detection algorithm is presented for 8-120 apple samples with a correlation coefficient ranging from 0.83 to 0.88 [79]. In addition, the blob detector neural network is demonstrated to detect the yield estimation for both oranges and apples with intersection over union of 81.3% for orange and 83.8% for apple [61]. As for disease identification, a simple linear iterative clustering algorithm is presented for 3624 foliar images with high classification rate of 98.34% for height between 1 to 2 m [98]. The classification rate is reduced for approximately 2% for each meter from the examined height within 1 to 16 m. Moreover, an improved GoogLeNet and Cifar10 models are established for 500 images of maize leaf disease with 4:1 ratio for training and validation which allows the system to have a diversity of sample conditions [100]. The average identification accuracy of GoogLeNet and Cifar10 models is recorded as high as 98.9% and 98.8%, respectively. Apart from bruise detection, yield estimation, and disease identification, algorithms are also shown in maturity evaluation and acquisition of crop segmentation. In maturity evaluation, both a Fuzzy model [62] and medium filter algorithm [85] are employed for 3108 images on banana samples and 100 images on tomato samples with an average identification rate of 93.11%, and within 89% to 98%, respectively. The Fuzzy model is useful in handling ambiguous information for the banana fruit maturity detection using red-green-blue (RGB) components. In the acquisition of crop segmentation, K-mean clustering algorithm is presented for a clustering of apple samples with target acquisition rate of 84% [80]. This algorithm is commonly used in image segmentation whereby crop segmentation can be precisely attained, even with the presence of stems and leaves in the captured images.
Applications of Spectroscopy Technique
The applications of spectroscopy technique in agriculture are presented in Table 6. The spectroscopy technique is widely applied to inspect internal qualities that are externally invisible. A sizeable amount of research has performed spectroscopy in the UV-VIS-IR region to identify the internal constituents of agricultural products such as pigment compound in apple [111], moisture content in mushroom [112], protein and sugar in potato [113], and caffeine in coffee [114] among others. Within 400 to 1000 nm, 678 nm is sensitive to low chlorophyll content thus the reflectance at 678 ± 30 nm is suggested for the monitoring of the early stage of ripening and the pigment content change with a maximum correlation of closely 0.6 [111]. On the other hand, 590 to 700 nm is recommended for the maturity detection in early stage for yellow colour apple fruits with maximum correlation from 0.7 to 0.9. In the verification of moisture content in mushroom, the spectral region from 600 to 2200 nm gives the lowest standard deviation of cross validation as 0.644% and maximum correlation factor of 0.951 among the investigated wavelengths from 402 to 2490 nm [112]. A high experimental repeatability is presented by a standard deviation of 0.677% and a maximum correlation factor of 0.947 for a separate set of mushrooms of a similar type and treatment. In the protein and sugar content identification in potato, a modified PLS regression model is applied to calculate the relationship between the spectrum and chemical properties of the calibrated samples [113]. Based on the measurement, the standard deviations for crude protein, glucose, fructose, sucrose and red sugar for the 120 potato samples are 0.2%, 0.073%, 0.068%, 0.068%, and 0.122%, respectively. Correspondingly, the squared correlation coefficient for the above five parameters are deduced as 0.96, 0.70, 0.89, 0.62, and 0.82, respectively. In a total of 665 tea leaf samples, NIRS and liquid chromatography is coupled to a diode arrayed detector to determine its content of caffeine [114]. Among 375 calibration sets and 250 validation sets for caffeine in the tea leaf samples, a standard deviation of 8.6 and 8.9, as well as high squared correlation coefficients of 0.97 are acquired for both calibration and validation sets though regression model.
The quality and freshness of fruits [115,116], vegetables [117], and meat [118,119] can be easily inspected using spectroscopy as well. For instance, two wavelengths within 600 to 904 nm of VIS-NIR spectrum are investigated by correlation analysis to discriminate brown core and sound pears [115].
Using eight brown core pears and 32 sound pears, the percentage of soluble solid content achieves a precision of 97.8% and 99% within a standard deviation of 0.5% and 1%, respectively. In addition, NIR spectrum and PLS regression model are used to detect the total anthocyanins content (TAC) and total phenolic compounds (TPC) in jambu fruits [116]. With a total of 50 jambu samples scanning from 1000 to 2400 nm, the correlation coefficients of TAC and TPC are deduced as 0.98 and 0.94, and strong ratio to performance of deviation as 5.19 and 3.27, respectively. Besides that, a 250 to 350 GHz radiation is found to be suitable to distinguish the defective and proper sugar beet seeds [117]. A python package scikit-learn algorithm is used to determine the threshold for these two types of seeds, with 80% detection for proper seed and 94% detection for defective seed. Therefore, the average detection rate of this algorithm is 87%. In addition, meat fraud is injected into bovine meat, aiming to increase the water holding capacity. This issue is characterized with attenuated total reflectance Fourier transform infrared spectrum and the supervision of the 55 meat fraud adulterated samples through PLS square discriminant analysis [118]. The analysis records a precise detection as high as 91% of the adulterated samples. Apart from meat fraud, the freshness of mackerel fish is characterized with auto-fluorescence spectroscopy and analyzed with fluorescence excitation emission matrices (EEM) [119]. The fluorescence EEM data and real freshness values are modelled with PLS regression and an algorithm is developed for this smart system as a predictor with squared correlation coefficient of 0.89.
Furthermore, chemical residues in harvested product [120] or even plantation soil [121][122][123] can be easily identified, leading to easy detection on contamination of agricultural product. Residual pesticides such as phosmet and thiabendazole in apples are analyzed with surface-enhanced Raman spectroscopy (SERS) coupled with gold nanoparticle [120]. The sensitivities for detectable concentration are 0.5 µg/g for phosmet and 0.1 µg/g for thiabendazole. The PLS regression is also used to correlate the SERS spectrum with the concentration of pesticide in apples with squared correlation coefficient of 0936 for phosmet and 0.959 for thiabendazole. In addition, the effect of drying temperature on the nitrogen detection in soils at four different temperatures of 25 • C for placement, 50, 80, and 95 • C for drying is modelled based on NIR sensor and three successive algorithms, which are multiple linear regression, PLS, and competitive adaptive reweighed square on the spectral information [122]. Based on the three soil samples, loess, calcium soil, and black soil show the correlation coefficients of 0.9721, 0.9588, and 0.9486, respectively at the optimum drying temperature of 80 • C. The detection of nitrogen in three types of soils is also alternatively performed in [123], with squared correlation coefficients of 0.95, 0.96 and 0.79 for loess, calcium soil and black soil using PLS regression model. The relatively lower squared correlation coefficient in black soil is due to the interference of high humus content and strong absorption of organic matter in black soil. Lastly, a point worth noting is that the dielectric and NMR spectroscopy are often adopted when the analysis involves more complex chemical compounds [124,125]. These complex chemical compounds include the vulcanization of natural rubber with sulfur-cured and peroxide-cured systems with different dynamics [124] and detection to changes in concentrations of pollutants in agriculture drainage such as heavy metal and heavy oxides [125]. In [124] a sulphur-cured system has features restricted segmented dynamics whereas peroxide-cured system has faster dynamics. In addition, network structure resulting from the vulcanization of both systems also influences the segmental dynamics of natural rubber. The peroxide-cured network is more homogeneous with spatial distribution of cross links than the sulphur-cured network with large inhomogeneity due to the presence of zinc oxide particles and the ionic interaction with the natural rubber chains. In [125], an X-ray fluorescence spectroscopy is employed to investigate the changes of pollutants in dried root and shoot plant parts at a temperature range from 30 to 90 • C. From the measurement, the concentration of pollutant is found to be higher in plant root than plant shoot through the analysis of frequency relaxation process via dielectric modulus measurement. Significantly, the removal of pollutants by plants will be enhanced upon subjecting them to a microwave heating power of 400 W for 30 min. Apart from the aforementioned applications, more applications of the spectroscopy technique in agricultural products with different methods and wavelengths are tabulated in Table 6.
Applications of Spectral Imaging Technique
As discussed earlier, the spectral imaging technique is a combination of both imaging and spectroscopy techniques. In the agriculture industry, both variations of hyperspectral and multispectral imaging are equally crucial, and some sample applications are compiled in Table 7. As observed from Table 7, hyperspectral imaging involves acquisition over a range of wavelengths while multispectral imaging involved acquisition at fewer selected wavelengths. The robustness of spectral imaging allows its usage in bruise detection [78,162,163], maturity evaluation [164][165][166][167][168], quality evaluation [169][170][171] and disease detection [172][173][174][175][176]. Internal attributes of samples [177][178][179] are easily acquired for analysis purposes as well.
In bruise detection of agriculture product, a machine vision system is integrated with optical filter at 740 and 950 nm to detect the bruise in rotating apple with a detection accuracy of 90 to 92% from 54 Pink Lady apples and 60 Ginger Gold apples [162]. In addition, the bruise detection in mushroom is carried out through a line scanning hyperspectral imaging instrument from 400 to 1000 nm with a spectroscopic resolution of 5 nm [163]. The PCA is applied to a set of data comprising of 50 normal and 50 bruise spectra, with standard deviation of 0.025 and 0.055, respectively.
For the maturity evaluation of agriculture products such as peach fruit, a CCD camera is employed at 450, 675 and 800 nm, whereas the fruit ripening is characterized with the increasing in intensity from a histogram with a ratio of red divide with infrared red (R/IR) [164]. The firmness of the peach fruit reduces when the reflectance at 675 nm is increased. An analysis of variance (ANOVA) is presented to access the R/IR clustering which has the highest reflectance at 675 nm, and higher Fisher value as a function of higher R/IR ratio. Apart from the detection of maturity for peach fruit, the maturity stage of strawberry is detected using a hyperspectral imaging system from 380 to 1030 nm and from 874 to 1734 nm [165]. According to the PCA, the optimal wavelengths are from 441.1 to 1013.97 nm and from 941.46 to 1578.13 nm with a classification accuracy of above 85%. Moreover, the maturity of tomato is detected by an electromagnetic spectrum with a continuous 257 bands from 396 to 736 nm [166] and discrete band of 530, 595, 630, and 850 nm using a tomato maturity predictive sensor [167]. Based on the LDA, the classification error is reduced from 51% to 19% [166] and achieves detection accuracy above 85% [167]. The ripening in banana fruit is also characterized with a compact imaging system and an UAV from 500 to 700 nm [168]. The detection is based on the reflectance spectrum whereby in ripe banana, the main element is carotenoid which absorbs less light at 650 nm band. On the other hand, a green banana with a greater amount of chlorophyll than ripe banana absorbs more light at 650 nm band.
For the quality evaluation of agriculture product, the firmness test for two types of apple fruits is conducted with a laser-based multispectral imaging prototype which captures and processes four spectral scattering images at a speed of two fruits per second [169]. The multilinear regression models are developed using to predict the firmness of those two apple types at 680, 880, 905, and 940 nm with a correlation coefficient of 0.86. The quality of grape berries is determined by the reflectance spectrum from a hyperspectral imaging system, whilst the high reflectance at 500 nm, 660 to 700 nm, and 840 nm denotes the chlorophyll content, red-coloured anthocyanin pigment, and sugar content, respectively [170]. A partial least square regression (PLSR) model is applied in order to determine the correlation between the spectral information and the physico-chemical indices. The titratable acidity of the green and black grapes shows a coefficient of determination of 0.95 and 0.82, as well as soluble solid content of 0.94 and 0.93 at pH value of 0.8 and 0.9, respectively. The root mean square error for this method is 0.06 for green grape and 0.25 for black grape. Apart from fruits, the quality of tea leaves is classified by a hyperspectral imaging sensor at 762, 793, and 838 nm, supported by SVM algorithm [171]. Within 700 samples comprising of 500 training samples and 200 prediction sets, the SVM algorithm generates a total classification rate of 98% the training sample and 95% for the prediction set, at result of optimal regularization parameter of 4.37349 and kernel parameter of 13.2131 in SVM model.
For the disease detection in agriculture product, a fruit sorting machine is used to detect the citrus canker at 730 and 830 nm with a bandpass filter installed in the scanning camera [172]. A real-time image processing and classification algorithm is developed based on a two-band ratio (R830/R730) approach, which achieves a detection accuracy of 95.3% for 360 citrus samples. Next, a shortwave infrared hyperspectral imaging system is used to detect the sour skin in onion based on the suitable reflectance spectrum from 1070 to 1400 nm [173]. Two image analysis approaches are utilized based on the log-ratio images at two optimal wavelengths of 1070 and 1400 nm. A global threshold of 0.45 is integrated to segregate sour onion skin infection areas from log-ratio images. With Fisher's discriminant analysis, the detection accuracy of 80% is achieved. The second image analysis approach is the incorporation of three parameters; max, contrast and homogeneity of the log-ratio images as the input features for the SVM. Subsequently, the Gaussian kernel generates higher detection accuracy as 87.14%. Apart from that, the tumorous chicken is detected by the combination of a CCD camera and imaging spectrograph from 420 to 850 nm [174]. Within the wavelength bands, the PCA select the three useful wavelengths of 465, 575, and 705 nm from the tumorous chicken image. Based on the images from 60 tumorous and 20 normal chicken, multispectral image analysis generates the ratio images, which are divided into ROI classified either as tumorous or normal chicken. The image features from ROI such as coefficient of variation, skewness and kurtosis are extracted as the input for the Fuzzy classifier, which generates the detection accuracy of 91% for normal chicken and 86% for tumorous chicken. To detect the nematodes in coffee cultivation, hyperspectral data is used in band simulation of the RapidEye sensor to determine the most sensitive spectral ranges for pathogen discrimination in coffee plants [176]. Multispectral classification identifies the spatial distribution of healthy, moderately infected, and severely infected coffee plants with an overall accuracy of 71%. Apart from the four main applications with the spectral imaging technique in agriculture products, more applications with different scanning methods and various wavelengths are tabulated in Table 7. [190,191] Thai jasmine rice Rice variety identification Multi. area scan 545, 575 [192] Wheat Fungus detection Hyper. area scan 1000-1600 [193] Wheat Damage detection Hyper. line scan 1000-2500 [194] Peanut
Photonics Techniques Implementation in Food Safety Inspection and Quality Control
Food safety inspection and quality control is important for ensuring the high quality of agriculture products. To meet this criterion, photonics techniques have been extensively implemented into numerous applications. For instance, clean drinking water is undeniably one of the most important elements to sustain the organisms' life. Contamination may happen when treated drinking water is travelling in the distribution system to the consumer, whilst the sensitivity to the inhibitor of contamination can be measured by the elevated dissolved organic matter (DOM) at the tap relative to the water leaving the treatment plant [216]. Across a biologically stable drinking water system, humic-like fluorescence (HLF) intensities of less than 2.2% relative standard deviation are measured after accounting for quenching by copper. In addition, a minor infiltration of a contaminant is detectable by sewage with a strong tryptophan-like fluorescence (TLF) signal thus validating the potential of DOM fluorescence in detecting the water quality changes in drinking water system. Moreover, fluorescence spectroscopy was demonstrated in evaluating the microbial quality of untreated drinking water through online monitoring [217]. The DOM peaks are targeted at excitation and emission wavelengths of 280 and 365 nm for TLF, as well as 280 and 450 nm for HLF. Both TLF and HLF are strongly correlated to micro-bacterial cells such as E. coli with a correlation coefficient of 0.71 to 0.77. In comparison to turbidity for E. coli with correlation coefficient of only 0.4 to 0.48, the DOM sensor appears to be a better indicator for micro-bacterial cells in untreated drinking water. Apart from the DOM sensor, an optical sensor was proposed to differentiate the particles in drinking water as either bacteria or abiotic particles with an accuracy of 90 ± 7% and 78 ± 14% for monotype and fix-type suspensions, respectively, based on a 3D image recognition and classification algorithm [218]. In addition, this optical sensor can detect micro-particles with minimum size of 0.77 µm. Significantly, the aforementioned optical sensors incorporating photonic techniques serve as an early warning for drinking water pollution.
Photonics and optics have also recently gained popularity in the quality inspection of food product. This is because food inspection in the production line needs to be carried out at fast speed and a very fast monitoring system is needed. Food inspection can become even more challenging when it is dealing with large quantities of sample moving very quickly on the conveyor belt. Therefore, high speed and high sensitivity optical system will be very suitable for the online monitoring and inspection of food product. For example, research work on UV-visible-NIR optical spectroscopy have been carried out extensively in the monitoring of extra virgin olive oil [219], honey [220], tea [221], dairy product [222] and alcoholic beverages [223]. However, as these works focus on wavebands below 1100 nm, the results and consistency of the conclusions may be easily affected by ambient lighting conditions and the change in color of the beverages or product. Therefore, more research work shall be carried out to characterize these food products in the NIR (>1100 nm) and MIR wavebands in order to obtain the optical "fingerprint" that correlates to the quality and food safety level of the product.
In addition, food preservative exceeding the allowable limit has been a critical issue in ensuring the health of the public. Butylated Hydroxytoluene (BHT) is commonly used as an antioxidant agent in canned food or bottled beverages. Several optical sensing techniques such as optical spectroscopy and fluorescence may be able to detect the concentration of BHT. BHT is also commonly known as 2,6-ditertiarybutyl-para-cresol (DBPC). Recently, Leong et al. [224] reported the detection of DBPC in transformer oil using optical spectroscopy at waveband near to 1403 nm. This opens up the opportunity of detecting the concentration of BHT in canned food or bottled beverages, leading to an online monitoring system that uses the optical spectroscopy method.
Due to the lack of attention paid during the preparation processes or due to the contamination of water and environment, hazardous residual materials are occasionally found in food. These hazardous materials include heavy metal, pesticides and antibiotics. Conventionally, the screening process or food safety inspections were carried out using laboratory-based equipment or measurement methods such as gas chromatography (GC), GC-mass spectrometry and high-performance liquid chromatography [225]. However, these methods only allow inspection based on sampling due to the high cost and long result waiting time. In this context, optical detection methods such as optical spectroscopy, Raman spectroscopy and fluorescence can be explored for their possible utilization in the online monitoring of food products in order to ensure that they are free of hazardous residual materials.
Photonics Techniques Implementation in Tropical Countries Agriculture
Blessed with wide spans of fertile soil, rich marine ecology, abundant rainfall and a tropical climate, tropical countries are exceptionally suited for a myriad of agricultural activities [226][227][228]. Agriculture activities boost the country's economy by supplying food sources and industrial raw materials. This sector also provides income to farmers, raising their living standards in rural areas. An example of tropical countries with active agricultural activities is Malaysia. Dating back to the early years following Malaysia's independence in 1957, the agriculture sector has been a signifficant driver towards socio-economic development in Malaysia. However, in the early 1980s, the growth of the agricultural field came to an abrupt halt due to the sharp decline in commodity prices, limited technical specialty, volatile rubber prices and lack of incentives [229,230]. Industrialization soon became the leading economic sector, with great focus directed towards manufacturing and services [230]. Fortunately, the agriculture sector is once again emphasized upon the Asian financial crisis in 1997, acting as a measure to minimize external economic shock by first strengthening the domestic economy [227,231]. Since then, agriculture has always been a major agenda item of Malaysian economic plans, with a recent target directed towards modernizing agriculture as drafted in the Eleventh Malaysia Plan [232]. To date, amidst industrial developments, Malaysia has approximately 4.06 million hectares of agricultural land, with 80% allocated for commercial crops such as palm oil, rubber, cocoa, coconut and pepper [229,233], while a portion of the remaining 20% was utilized for the cultivation of agro-food crops [226]. These remarkable statistics have validated the potential of a tropical country to develop its agricultural sector. Apart from Malaysia, other tropical countries such as Indonesia and Thailand are also actively involved in agriculture activities. The following sections will discuss some of the agricultural crops in tropical countries in which optics and photonic techniques can be easily integrated for automated plantation management, yield increment, quality inspection and disease control.
Implementation in Palm Oil-Related Activities
Palm oil is an extremely valuable commercial crop in tropical countries. Palm oil, which is extracted from oil palm, is often used as raw materials for the production of biofuel, biofertilizers, oleochemicals, biomass products, nutraceuticals and pharmaceuticals. In fact, tropical countries are among the global leaders in the palm oil industry [234]. The implementation of optics and photonic techniques in palm oil-related activities will maintain the competitive power of the tropical countries in the field and help to reap the associated economic benefits.
The implementation of optics and photonics techniques in oil palm related activities can start from the development of agriculture robots. The development of an agriculture robot involves the implementation of the imaging technique in its operation. Spatial and color information attained by the agriculture robot through the imaging technique will greatly improve the efficiency of palm oil plantation management. Automated palm oil fruit harvest is potentially applicable by pinpointing the fruit position as presented in [72,74,75] for other crops. Besides, automated weed detection and removal [73] as well as automated fertilizing can be performed using the developed agriculture robot.
In addition, palm oil quality is governed by fatty acid, moisture and peroxide contents. Microbial or oxidation reactions that take place during the storage of oil palm fruit may modify these contents, resulting in a depreciation of palm oil quality [235]. Under common operations, palm oil plantations are usually distanced further away from refinery factories. Bulk transport of palm fruit upon reaching the necessary processing quota is often practiced for cost savings. As a result, palm fruits that have been harvested earlier will be stored in dedicated storage spaces. The time difference between harvesting and processing greatly increases the risk for microbial or oxidation reaction to take place. In this scenario, spectroscopy or spectral imaging can be implemented in the palm oil extraction stage to perform oil quality segregation. This will greatly prevent contamination of low-quality palm oil in further downstream processes, promoting process efficiency and increasing palm oil yield.
Another area in which optics and photonics techniques may be applied for oil palm activities is disease detection. The most devastating diseases that attack palm oil plantations in South East Asia are basal stem rot (BSR) and upper stem rot (USR). These diseases result in certain death of oil palms if not controlled effectively, resulting in yield loss and disrupting the plantation cycle. These fatal diseases are identified to be caused by the Ganoderma boninense (G. boninense) fungus. However, the identification of the root cause of these diseases is still insufficient as they cannot be controlled even with the slightest delay in infection detection [236]. In this area, spectral imaging presents itself as one of the possible alternatives to perform early detection of the G. boninense fungus [237]. Samples of suspicious fungi in the palm oil plantation can be simultaneously collected and analyzed to identify the presence of disease-causing G. boninense. From here, preventive measures can be effectively performed to curb any possible disease spreading.
Implementation in Natural Rubber Related Activities
Natural rubber is an important commodity that finds it place in the manufacturing of various household, industrial and medical products. Rubber tree plantations have been widely established in the fertile soils of tropical countries. The usage of optics and photonic techniques will again prove to be beneficial in this area.
The simplest idea will again start from the usage of agriculture robots during the plantation stage. In the context of rubber tree plantations, the imaging technique will provide visual guidance for the agriculture robots to perform the scheduled collection of field latex. The usage of these robots will gradually replace manual latex collection done by rubber tappers. This approach will address the decline in manpower to maintain rubber tree plantations.
Meanwhile, the spectroscopy technique can be utilized in the later rubber processing stages. The first application would be rubber quality grading. For instance, cup lump raw rubber, which is an important material in tires, seal strips, conveyor belts and other moulded rubber products, can be graded by using VIS-NIR spectroscopy to inspect the moisture content of the rubber. This spectroscopic approach is fast, accurate and more reliable compared to manual inspection through sight and touch [155]. Similarly, the protein and lipid contents in natural rubber can be detected through NIR-MIR spectroscopy to enable grading [152]. Lastly, spectroscopy variations, such as NIR-MIR, Raman, dielectric or NMR, can be opted to study the structure and properties of rubber during vulcanization. Such studies allow the analysis and selection of accelerators, activators and retarders, leading to improved characteristics in the vulcanized rubber and an optimized vulcanizing process [124,153].
Implementation in Agro-Food Crops Related Activities
It is important to increase food production and achieve a self-sufficiency level (SSL) for a growing country to become an advanced country. Currently, the agro-food crops in tropical countries comprise of grains, organic fruits and vegetables, herbs and spices, livestock and fisheries [232,234]. By referring to some of the applications stated in Sections 3.1-3.3, optics and photonic techniques can once again improve the overall quality and yield of these crops.
Starting from grains such as rice and corn, crop harvest [72] and weed removal [73] can be easily performed by agriculture robots with imaging capabilities. Thermal imaging can be conducted to evaluate water stress in crops for irrigation control [97]. Moreover, the development of mobile phone application to perform color-based identification of nitrogen content in rice and corn plant is another interesting idea. The usage of such applications promotes the portable and on-site analysis of fertilizer requirements in crop fields [64].
At the same time, all three optics and photonic techniques discussed earlier can be fully utilized to inspect the harvested organic fruits and vegetables for quality evaluation. For instance, imaging in either VIS or IR region is useful in detecting external damage or bruises in mangosteens, wax jambus, cherry tomatoes and more. Spectroscopy may be performed as well to inspect internal features or maturity of fruits and vegetables. Not least, spectral imaging may be considered when spatial and spectral information are required simultaneously for quality evaluation. Meanwhile, the quality inspection of meat products, such as chicken, beef, lamb, and fish among others, is strongly preferred to be performed using spectroscopy or spectral imaging. These two techniques are suitable for identifying the microbial spoilage of meat products due to their ability to obtain spectral information. With the integration and application of optics and photonics in the agriculture industry, it is anticipated that the agricultural products in the tropical countries will meet the public expectation of higher food quality.
Possible Challenges
The prevailing research challenges of integrating optics and photonics techniques into the agriculture field are the reliability issue of the laser source and sensor, effect of the ambient environmental condition into optics system, and expensive semiconductor materials at operating wavelength from short to mid-IR range. First and foremost, the illumination intensity of the laser and the sensitivity of the sensor may change over time, which leads to the need for recalibration of the system. Therefore, more research is required in terms of the design and fabrication of a more reliable laser source, sensor and optical detector. In addition, the effect of the ambient condition such as humidity, surrounding temperature, and dust particles could be a hindrance in ensuring consistent results obtained from the optical system. Hence, research into the minimization of these effects on the optical system is significant to improve the system performance such as higher sensitivity, lower systematic error and maintenance rate. Moreover, silicon is well-known for its optimum wavelength operation below 1000 nm. From short to mid-IR range, examples of more viable semiconductor materials are gallium antimonide and indium gallium arsenide. The investigation in terms of generating a higher efficiency using these materials for a cost-effective solution creates the research opportunities for further exploration in both simulation and experimental works.
Apart from the research challenges, the main challenge in introducing the discussed optics and photonic techniques into the field of agriculture in tropical countries would be gaining the acceptance of farmers, fisherman and smallholders. The introduction of modern technology and new agriculture practices often raises concerns surrounding their technical and economic feasibilities. Farm and plantation owners will prefer traditional agriculture practices as newly introduced technologies are often regarded to be more suited to a controlled laboratory environment. In this scenario, technology vendors should ensure that complete field testing has been done in the environment where the technology will be introduced. A probationary period may also be set to allow owners to try out and experience the benefits brought forth by the proposed technologies.
The next challenge would be on financial limitations. In general, the cost to fully implement optics and photonics techniques in existing agriculture activities may be a burden to the owners, especially those involving sophisticated optical tools. This deterring factor may be mitigated if financial aids are provided to the owners. In this case, the government of tropical countries should set the right path by providing funds to the owners through attractive policies. For instance, a loan policy of flexible repayment based on harvest cycles is more attractive compared to one of fixed term financing since owners are now presented with flexible loans [232].
Lastly, another challenge lies with the need of technical support. When introducing the optics and photonic techniques, technical training should be provided to farm and plantation workers in order to familiarize them with the operations of new tools. At the same time, advisory and technical services should be easily available in case the agriculture tools experience downtime or require scheduled maintenance.
Conclusions
In conclusion, optics and photonics exhibit great benefits if they are integrated into the agricultural industry. A complete knowledge of the behaviors and properties of light upon light-material interaction allows the quantitative and qualitative analysis of agriculture products. In general, optics and photonic techniques for agricultural purposes can be categorized into imaging, spectroscopy and spectral imaging techniques. The imaging technique is effective in collecting spatial, color and thermal information, whereas the spectroscopy technique is essential for collecting spectral information. Meanwhile, spectral imaging is a combination of both imaging and spectroscopy techniques, allowing the collection of a complete data set. These three optics and photonic techniques have been utilized in agriculture categories such as fruits, vegetables, grain, meat, dairy produce, oil, beverages, and commercial crops, as well as farm and plantation management. These works can be referred to and emulated in the agriculture industry of tropical countries, especially in agriculture activities related to oil palm, rubber and agro-food crops. However, challenges in terms of public acceptance, finance and technical support should be overcome before achieving a complete integration of optics and photonics techniques in the agriculture industry.
Thus, the key contribution of this study is the comprehensive analysis of different optics and photonics systems in agricultural applications to provide a detail idea of the advanced techniques and their future deployment in agriculture cultivation and harvesting. The review has proposed important and selective suggestions for the further technological development of optics and photonics in future agricultural applications:
•
The incorporation of optical sensors into photonics detection techniques that serve as an early warning for drinking water pollution.
•
The characterization of canned food or bottled beverages in the NIR (>1100 nm) and MIR wavebands for their optical "fingerprint" that correlates to the quality and food safety level of the product, such as preservatives concentration.
•
The characterization on hazardous residual materials in food using optical spectroscopy, Raman spectroscopy and fluorescence.
•
The implementation of an agricultural robot to perform better palm oil plantation management, scheduled collection of field latex and weed removal.
•
The spectral imaging provides early detection of disease-causing G. boninense in the oil palm. • Spectroscopy provides moisture content inspection, protein and lipid content detection, as well as improving the rubber vulcanizing process.
•
The imaging technique detects external damage or bruises on organic fruits and vegetables. | 20,500.2 | 2019-05-01T00:00:00.000 | [
"Agricultural and Food Sciences",
"Engineering",
"Physics"
] |
Multi-Angle Lipreading with Angle Classification-Based Feature Extraction and Its Application to Audio-Visual Speech Recognition
: Recently, automatic speech recognition (ASR) and visual speech recognition (VSR) have been widely researched owing to development in deep learning. Most VSR research works focus only on frontal face images. However, assuming real scenes, it is obvious that a VSR system should correctly recognize spoken contents from not only frontal but also diagonal or profile faces. In this paper, we propose a novel VSR method that is applicable to faces taken at any angle. Firstly, view classification is carried out to estimate face angles. Based on the results, feature extraction is then conducted using the best combination of pre-trained feature extraction models. Next, lipreading is carried out using the features. We also developed audio-visual speech recognition (AVSR) using the VSR in addition to conventional ASR. Audio results were obtained from ASR, followed by incorporating audio and visual results in a decision fusion manner. We evaluated our methods using OuluVS2, a multi-angle audio-visual database. We then confirmed that our approach achieved the best performance among conventional VSR schemes in a phrase classification task. In addition, we found that our AVSR results are better than ASR and VSR results.
Introduction
Recently, automatic speech recognition (ASR) has been confirmed to have high recognition performance by using deep learning (DL), an attractive artificial intelligence technology, and is used in various scenarios, such as voice input for mobile phones and car navigation systems. However, there is a problem that speech waveforms are degraded by audio noise in real environments, reducing the accuracy of speech recognition. In order to overcome this issue, we need to develop robust ASR systems against any audio noise. One of these ASR systems applicable in noisy environments is audio visual speech recognition (AVSR, also known as multi-modal speech recognition), which employs ASR frameworks with visual speech recognition (VSR, also known as lipreading). VSR uses lip images which are not affected by audio noise and estimates what a subject uttered only from a temporal sequence of lip images. VSR and AVSR have a potential to be applied in various practical applications such as automatic conference minute generation and human interfaces on smartphones. Owing to state-of-the-art DL technology, recently, we have achieved high performance of VSR. However, VSR still has several problems when we employ the technique in real-world scenes; for example, most VSR studies have only considered frontal faces, but VSR technology for non-frontal views is also essential for real applications. In other words, assuming real scenes, a speaker does not always face a camera, such as smart device or tablet device, in a VSR or an AVSR system. We thus have been developing multi-angle VSR architecture which enables us to perform VSR when not only frontal lip images but also non-frontal lip images are observed.
There are two main approaches for multi-angle VSR. The first method is to build a VSR model using training lip images captured at several angles. The second approach is to convert non-frontal lip images to frontal ones and apply the conventional frontal VSR technique. In this paper, we focus on the first approach, and propose a feature integrationbased multi-angle VSR system using DL, particularly 3D convolutional neural networks (CNNs), that are one kind of deep neural networks (DNNs). Based on most conventional multi-angle VSR studies, it is necessary to estimate at which angle lip images are captured, to choose a suitable angle-specific VSR model. However, if the system fails to estimate the right angle, the recognition performance drastically decreases. We need to build a VSR technique that can be applied to real scenes where it is difficult to estimate the accurate lip angle.
Therefore, we employ a new multi-angle VSR method, in which all angle-specific VSR models are trained using images at different angles. Our multi-angle VSR method consists of three parts: a view classification part, a feature extraction part and a recognition part. Assume that we have a sequence of lip images to be recognized. Firstly, in the view classification part, we prepare a common 2D CNN that estimates the angle of the input image (see Section 3.1.1). The model is then applied to each image in the sequence, followed by determining the angle which has the majority in the estimation. Secondly, in the feature extraction part, we build 3D CNN models for possible combinations of angle-specific training data sets (see Section 3.1.2). Based on the angle obtained in the first part, we choose the best models and extract features from the models. In the last integration part, we concatenate these features, followed by recognition by means of a fully connected (FC) neural network (see Section 3.1.3) In addition, we perform a decision fusion-based AVSR employing our proposed multi-angle VSR.
We conducted evaluation experiments using the open data set OuluVS2, in which subjects were captured simultaneously at five angles in addition to speech data. The experimental results show that our proposed method can improve VSR accuracy much more than conventional schemes on average, and achieve significant AVSR accuracy in noisy environments. In addition, we confirm that our proposed method is sufficiently robust against view classification errors, because, in the second part, we simultaneously employ several models built using multi-angle training data.
The rest of this paper is organized as follows. In Section 2, we briefly review related works on multi-angle VSR. Section 3 introduces our method. The experimental setup, results and discussion are described in Section 4. Finally, Section 5 concludes this paper.
Related Work
Recently, many researchers have proposed deep learning-based AVSR and VSR schemes . As mentioned, most conventional VSR research has focused on frontal face images, assuming that VSR systems are in front of speakers, since there are only a few data sets available with multi-angle faces. Here, we introduce several lipreading works focusing not only on frontal but also diagonal and profile images. To develop these schemes, we need a research corpus. One of the public multi-angle VSR data sets is OuluVS2 [22].
An early work of multi-angle lipreading is [1], where a system was trained using either frontal (0 • ) or profile (90 • ) faces. According to the experimental results, the frontal view showed a lower word error rate (WER) than the profile view. In [2], the authors built a multi-angle system investigating a frontal (0 • ) view, a left profile (90 • ) view and a right profile (−90 • ) view. They reported significantly better performance when using the frontal view than the others. Saitoh et al. proposed a novel sequence image representation method called concatenated frame image (CFI) [3]. Two types of data augmentation methods for CFI, and a framework of a CFI-based CNN, were tested. Bauman et al. indicated that human lipreaders tend to have higher performance when slightly angled faces are available, presumably because of the visibility of lip protrusion and rounding [4]. In [5], the active appearance model (AAM) was utilized for feature extraction at five angles, and lipreading was examined on a view-dependent system, as well as on a view-independent system using a regression method in a feature space. As a result, the view-dependent system performed the best performance at 30 • in all tests. Zimmermann et al. used principal component analysis (PCA)-based convolutional networks together with Long short-term memories (LSTMs), one of the DL models, in addition to a conventional speech recognition model, hidden Markov models (HMMs) with Gaussian mixture models (GMMs) [6]. They aimed at combining multiple views by employing these techniques. They finally confirmed that the highest performance was obtained at 30 • . Anina et al. stated that the highest accuracy was achieved at 60 • in their experiments [22]. Kumar et al. showed that profile-view lipreading provides significantly lower WERs than frontal-view lipreading [7].
There is another strategy to conduct transformation to images or incorporate several views with DL technology. There is one work [8] that involved converting faces viewed from various directions to frontal faces using AAMs. The experimental results showed that recognition accuracy was improved even when the face direction changed about 30 • relative to a frontal view. In [9], the authors proposed a scheme called "View2View" using an encoder-decoder model based on CNNs. The method transformed non-frontal mouth region images into frontal ones. Their results showed that the view-mapping system worked well for VSR and AVSR. Estellers et al. introduced a pose normalization technique and performed speech recognition from multiple views by generating virtual frontal views from non-frontal images [10]. In [11], Petridis et al. proposed an end-to-end multi-view lipreading system based on bidirectional LSTM networks. This model simultaneously extracted features directly from the pixels and performed visual speech classification from multi-angle views. The experimental results demonstrated that the combination of frontal and profile views improved accuracy over the frontal view. Zimmermann et al. also proposed another decision fusion-based lipreading model [12]; they extracted features through a PCA-based convolutional neural network, LSTM network and GMM-HMM scheme. The decision fusion succeeded by combining Viterbi paths. In [13], Sahrawat et al. extended a hybrid attention-based connectionist temporal classification system with view-temporal attention to perform multi-angle lipreading. Lee et al. trained an end-to-end CNN-LSTM model [14].
Many studies have been conducted focusing on AVSR. In this paper, we would like to introduce a couple of state-of-the-art works. An AVSR system based on a recurrent neural network transducer architecture was built in [15]. The authors evaluated the system using the LRS3-TED data set, achieving high performance. In [16], the authors proposed a multimodal attention-based method for AVSR, which could automatically learn fused representations from both modalities based on their importance. They employed sequence-to-sequence architectures, and confirmed high recognition performance under both acoustically clean and noisy conditions. Another AVSR system using a transformerbased architecture was proposed in [17]. The experimental results show that on the How2 data set, the system improved word error rate relatively over sub-word prediction models. In [18], we proposed an AVSR method based on deep canonical correlation analysis (DCCA). DCCA consequently generates projections from two modalities into one common space, so that the correlation of projected vectors could be maximized. We thus employed DCCA techniques with audio and visual modalities to enhance the robustness of ASR. As a result, we confirmed that DCCA features of each modality can be improved compared to the original features, and better ASR results in various noisy environments can be obtained.
Although we can find a lot of VSR and AVSR methods, there are only a few works combining ASR and multi-angle VSR to accomplish angle-invariant AVSR. One of them is [19], where the authors proposed an early fusion-based AVSR method using bidirectional LSTMs. Similar to their past work [11], the authors put lip images at various angles and corresponding audio signals into the bidirectional LSTM models.
Methodology
Our proposed multi-angle VSR method consists of three parts: a view classification part, a feature extraction part and a recognition part. Figure 1 depicts the architecture of our AVSR approach, including ASR and the VSR model. In this section, we describe each part of our multi-angle VSR scheme followed by ASR and AVSR frameworks.
Multi-Angle VSR
VSR accepts a temporal sequence of lip images to recognize what a subject utters according to the given images. Assuming real scenes, it is not guaranteed that a speaker is strictly facing a VSR system. One way to deal with this problem is to prepare several models, each of which corresponds to a certain angle, estimate at which angle face images are captured and apply a corresponding angle-specific model.
View Classification
In the view classification part, we at first estimate at which angle face images were recorded among the following five candidates in this work: 0 • , 30 • , 45 • , 60 • and 90 • . The estimation was carried out for each lip image in one sequence, using the 2D CNN model illustrated in Figure 2. The 2D CNN model employs a simple and common architecture; convolutional and pooling layers are repeatedly applied followed by FC layers, to obtain a classification result. After processing the above step for all the input images, we determine the angle which is the most often chosen.
Feature Extraction
Before conducting feature extraction, we prepare 3D CNN pre-recognition models for all possible combinations of the above five angles, i.e., models each trained only using images obtained from a single angle, such as a model from frontal images and a model from 30 • images, as well as models each built using data of several angles, such as a model trained using both 0 • and 30 • data and a model using all face images. An architecture of our 3D CNN-based VSR models is shown in Figure 3. The last layer has 20 outputs, each of which corresponds to one class in our recognition task. As a result, we build 31 models in this case (∑ 5 i=1 5 C i = 5 + 10 + 10 + 5 + 1 = 31), as shown in Table 1. Table 1 also indicates preliminary VSR results: recognition accuracy to validation data at a certain angle, using a certain model chosen among those 31 models. For example, if we adopt a 30 • model for 60 • data, the accuracy is 87.55%. architecture of our 3D CNN-based VSR models is shown in Fig. 3. The last layer has 20 169 outputs, each which corresponds to one class in our recognition task. As a result, we 170 build 31 models in this case (∑ 5 i=1 5 C i = 5 + 10 + 10 + 5 + 1 = 31), shown in Table 1. According to the angle obtained in the view classification part, we select the most reliable three models for the estimated angle, which are shown in bold in Table 1. For instance, we adopt (1) "0 • + 30 • + 45 • ", (2) "0 • + 30 • + 45 • + 60 • " and (3) "0 • + 30 • + 45 • + 90 • " models for 45 • data. In other words, we determine suitable angle combination patterns of training data for the estimated angle. We then utilize those models as feature extractors; we remove the last layer, resulting in a new output layer generating a 48-dimensional feature vector, as indicated in Figure 3. Finally, we obtain three 48-dimensional vectors from this part.
171
This strategy has two advantages. First, as shown in Table 1, models trained using data of several angles have relatively higher performance than those trained using single angle data. This result motivates us to choose such models for multi-angle data. Second, even if the view classification fails, it is still expected to obtain high performance by our scheme; for instance, in the case where a 30 • sequence is misclassified as 45 • , the above models (1)∼(3) are used for feature extraction, all in which 30 • data are also used in model training. There is another reason to encourage us to choose this framework. The model trained using all data, indicated in the bottom row in Table 1, achieved good performance. On the other hand, there exists a better model in all the angle cases. This suggests using only the model with all data is not the best solution. Hence, for each angle, we prepare several models trained using multi-angle data and utilize them as feature extractors.
Recognition
In the integration part, firstly, we integrate those 48-dimensional features extracted from three angle-specific models, by simply concatenating them. Thereafter, we conduct recognition using two FC layers (48 × 3 → 48 → 20). Here, we apply a 50% dropout between the FC layers. In our ASR framework, we extract 13 mel-frequency cepstrum coefficients (MFCCs) in addition to 13 ∆MFCCs and 13 ∆∆MFCCs from audio waveforms with a frame length of 25 msec and a frame shift of 10 msec [23][24][25][26]. The MFCC is the most commonly used feature in the speech recognition field in addition to ∆MFCCs and ∆∆MFCCs, which are first and second derivatives, respectively. As a result, we obtain a 39-dimensional acoustic vector.
In the acoustic modality, there are many frameworks and a lot of features, e.g., [27,28]. We should carefully choose an audio processing scheme based on performance and theoretical perspectives. For instance, mel-frequency spectrograms are commonly used for CNN-based speech recognition. In this study, we first conduct preliminary experiments to measure the accuracy when using mel-frequency spectrograms or MFCCs. The size of the spectrograms is 96 × 128. Because using MFCCs with CNNs achieves better performance, we choose this framework. Note that we need to investigate which acoustic processing methods and features are the most suitable for the other tasks.
Recognition
After computing MFCCs from consecutive frames, we apply a 2D CNN-based model for recognition, which is illustrated in Figure 4. Similar to the VSR model, we finally obtain an audio result including a probability for each class.
AVSR
Firstly, a sequence of lip images is added to the VSR model, while corresponding speech data are given to the ASR model. As mentioned in detail later, we adopt the corpus OuluVS2, in which the task is to estimate which sentence is spoken. Therefore, for each class, we obtain a probability from ASR results and another one from VSR. These probabilities are integrated in a decision fusion manner. Let us denote conditional probabilities of class c from ASR and VSR models by P A (x A |c) and P V (x V |c), respectively. Here, x A indicates an audio input representation, and x V means the corresponding image vector. We then obtain an audio-visual probability P AV (x A , x V |c) as: In this work, we simply fix α = 0.5.
Experiments
In order to examine the effectiveness of our VSR scheme as well as AVSR framework, we carry out recognition experiments.
OuluVS2
We choose the OuluVS2 corpus to evaluate our scheme. The database contains 10 short phrases, 10 digits sequences and 10 TIMIT sentences uttered by 52 speakers. The corpus includes face images captured by five cameras simultaneously at 0 • (frontal), 30 • , 45 • , 60 • and 90 • (profile) angles. In this study, we adopt the phrase data and digit data, uttered three times by each speaker. In our experiment, the data spoken by 52 speakers are divided into training data by 35 speakers (speaker ID:1-36), validation data by 5 speakers (speaker ID: 37-41) and testing data by 12 speakers (speaker ID: 42-53). Note that the speaker ID: 29 is missing. We conduct the same data split as previous works, such as [3,6,14], for a fair comparison. We also check whether the data split is appropriate by changing the different split settings, and confirm that using the data sets gives us fair results. The phrases are as follows: "Excuse me", "Goodbye", "Hello", "How are you", "Nice to meet you", "See you", "I am sorry", "Thank you", "Have a good time", "You are welcome". Each digit utterance consists of 10 digits randomly chosen. Note that, since we use a part of this corpus to enhance model training data, the task in this work is a 10-class classification for phrase utterances.
DEMAND
We select another database, DEMAND [29], as a noise corpus. This corpus consists of six primary categories, each of which has three environments. Four of those primary categories are for closed spaces: Domestic, Office, Public and Transportation. The remaining two categories are recorded outdoors: Nature and Street. In this study, we add some of those noises to build audio training data.
CENSREC-1-AV
CENSREC-1-AV [30] is a Japanese audio-visual corpus for noisy multi-modal speech recognition. CENSREC-1-AV provides audio utterances, lip images and audio noise. In this study, we utilize the audio noise, i.e., interior car noises recorded on city roads and expressways, to obtain acoustically noisy testing data.
Experimental Setup
We evaluate a model by utterance-level accuracy: where H and N are the number of correctly recognized utterances and the total number of utterances, respectively. In addition, we also evaluate our model performance by the F1 score. An F1 score can be computed as: where Precision = T P T P + F P , Recall = T P T P + F N In Equation (4), T P is the number of correctly classified utterances. F P and F N indicate false positives and false negatives, respectively. We calculate the score in each class.
Since DNN-based model performance slightly varies depending on the probabilistic gradient descent algorithm, which is a common model training approach, we repeat the same experiment three times and the mean accuracy is calculated. In terms of DNN hyperparameters, we choose a cross-entropy function as a loss function and Adam as an optimizer. Batch size, epochs and learning rate are set to 32, 50 and 0.001, respectively. We carry out our experiments using NVIDIA GEFORCE RTX 2080 Ti.
Preprocessing
The OuluVS2 data set includes extracted lip images, however, the image size is not consistent. In order to apply DNNs, we resize all images to 64 × 64. Based on our preliminary experiments with different image sizes, considering classification accuracy and computational cost, we use the image size of 64 × 64. Furthermore, we normalize a frame length to 64; if the length is less than 64 we conduct upsampling, otherwise we suppress some frames. In addition, we convert all color images to gray-scale ones. Similar to visual frames, we normalize the audio frame length to 115; if the length is less than 115 we add last frame, otherwise up to 115 frames are used.
In the OuluVS2 corpus, there are 1050 (35 speakers × 10 utterance × 3 times) sentences available. However, the data size is not enough for DNN model training. To compensate for the lack of training data, we apply data augmentation in the audio and visual modalities. In the audio modality, we add acoustic noises in DEMAND to the original utterance data. The details, including noise type and signal-to-noise Ratio (SNR) conditions, are shown in Table 2. In the visual modality, we train our VSR models using not only phrase data but also digit sequence data based on our previous work [31]. First of all, we investigated view classification performance. View classification results for the test data are shown in Table 3. The whole accuracy of view classification was 91.39%. Focusing on the results for each angle, classification for frontal and profile views was fully successful. On the other hand, misclassification was found in the diagonal views, particularly at 45 • . In conclusion, the performance of our view classification was acceptable. However, the last fact also indicates that it is required for the following VSR models to carry out recognition successfully even for the miscategorized sequences. Table 4. We firstly tested our models with and without view classification. Our method with the view classification part achieved almost the same or better performance, compared to ours without the classification, in which the classification result was correctly given. This indicates our feature extraction and recognition strategy can perform well. Next, we compared our approach with conventional methods. Focusing on the average of recognition accuracy, our proposed method achieved the highest accuracy regardless of the presence or absence of the view classification part. It is interesting that at 45 • we found much more improvement than in the other conditions, and even the view classification performance was insufficient. Since 45 • data were used as training data in the neighboring 30 • and 60 • conditions, we might obtain such an improvement even if the view classification fails. We also found that our method was particularly effective in the medium-angle (30 • , 45 • and 60 • ) conditions, while the end-to-end system had higher accuracy for frontal and profile images. Figure 5 indicates F1 scores for each angle. Among all the angles, it is found that shorter utterances were relatively hard to classify, because there were fewer cues for recognition. Table 5 shows recognition accuracy of our ASR, VSR and AVSR methods in various noise environments. Note that, because the task was a 10-class classification, the accuracy in noisy environments tended to be higher compared to large-vocabulary speech recognition. The VSR accuracy was stable and unrelated to SNR since visual information is not affected by noise. As is already known, the results of VSR were lower than those of ASR in all the SNRs, because audio features are more effective and informative than visual ones. Among the models, AVSR achieved the best accuracy in all the conditions. In particular, at 0 dB, where the effect of noise was the largest, the performance was improved by 3% for city road noise and by 2.3% for expressway noise compared to ASR results. Even in the case of 20 dB, where the effect of noise was quite small, the accuracy was slightly improved. As mentioned, we employed the decision fusion strategy, which is the simplest integration method. Similar to the ensemble approach, we believe our decision fusion method could successfully integrate ASR and VSR results, which had different recognition errors.
Conclusions
In this paper, we proposed a multi-angle VSR system in which feature extraction was conducted using angle-specific models based on view classification results, followed by feature integration and VSR. We also proposed a decision fusion-based AVSR. We employed DNNs in our system, to perform view classification, feature extraction and recognition. The advantages of our method are choosing appropriate feature extraction models based on angle classification results, reducing the negative impact of misclassification, and incorporating ASR and VSR results efficiently. Evaluation experiments were conducted using the multi-view corpus OuluVS2. Then, we found our scheme could work well compared to past works, and we clarified the effectiveness of view classification and feature extraction from pre-trained angle-specific models. Moreover, we found that our AVSR method is superior to ASR and VSR because our decision fusion method could successfully integrate ASR and VSR results.
As our future work, we are planning to conduct experiments using different angle settings and other tasks. The implementation of this framework for real applications is also expected. In addition, because there are some research works investigating spectrograms instead of MFCCs, we will try to employ spectrograms as acoustic input. Finally, we will explore the suitable model architecture and its physical meaning for feature extraction. Funding: This research received no external funding Data Availability Statement: The databases used in this article are OuluVS2, DEMAND and CENSREC-1-AV. For details, please refer to [22], [29] and [30], respectively.
Conflicts of Interest:
The authors declare no conflict of interest. | 6,211.8 | 2021-07-15T00:00:00.000 | [
"Computer Science"
] |
Resistance to Noise of Non-linear Observers in Canonical Form Application to a Sludge Activation Model
The aim of this study was to increase the resistance to noise of an observer of a non-linear MISO system transformed into canonical regulation form of order n. For this, the principle idea was to add n observers on the output equations of the main observer. By adjusting the time scale of the output observers, the resistance to noise of the final estimates is considerably increased. The proposed method is illustrated by model simulations based on a non-linear Sludge Activation Model (SAM)
Introduction
State observers have been intensely exploited since (Luenberger, 1966), to model, control or identify linear and non-linear systems, including the studies of (Krener & Isidori, 1983;Zheng, Boutat, & Barbot, 2009) relating to non-linear systems transformable into a canonical form.The key idea in such approaches is to produce approximate measures of non-linearity of order 1, as in Extended Luenberger Observers (ELO) (Ciccarella, Mora, & Germani, 1993).Approximations of nonlinearities in the canonical form (which results in ELO) have already been suggested (Bestle & Zeitz, 1983), and this approach can be extended to higher order approximations (Röbenack & Lynch, 2004).An observer using a partial nonlinear observer canonical form (POCF) (Röbenack & Lynch, 2006) has weaker observability and integrability existence conditions than the well-established non-linear observer canonical form (OCF). Non-linear sliding mode observers use a quasi-Newtonian approach, applied after pseudo-derivations of the output signal (Efimov & Fridman, 2011).State observers using Extended Kalman Filters (EKF) provide another method of transforming non-linear systems (Boker & Khalil, 2013), (Rauh, Butt, & Aschemann, 2013).Finding an appropriate method for parameter synthesis remains one of the major difficulties with state observers for non-linear systems.(Tornambè, 1992), (Mobki, Sadeghia, & Rezazadehb, 2015) proposed high-gain state observers to deal with this problem.High-gain state observers reduce observation errors for a range of predetermined amplitudes or fluctuations by making the observations independent of parameters.The weak point of this method is its sensitivity to noise and uncertainty.In network identification and encryption, observers with delays are used to synchronize chaotic oscillators, as shown in several studies (Ibrir, 2009;Martínez-Guerra, et al., 2011).Noise and uncertainty are not critical factors in such a context.This can be very different in the case of industrial processes, as shown in a recent study (Bodizs, 2011), where the performances of observers using ELO, EKF or Integrated Kalman Filters (IKF) are compared.The influence of noise and uncertainty on these observer types was emphasized, with more reliable results produced by ELO observers, which permit the exact state reconstruction of highly perturbed systems.For PI and ELO observer classes, (Söffker, et al., 2002) demonstrated a compensation effect on measurement errors ; (Khalifa & Mabrouk, 2015) addressed the problem of uncertainty of non-linear models.One way of overcoming the problem of parametric uncertainty is to use adaptive observers (Tyukina, et al., 2013;Farza, et al., 2014) in the particular case where the measurements are only available at discrete instants and have disturbances.Another approach (Mazenc & Dinh, 2014;Thabet, et al., 2014) consists of defining interval observers.Modeling observer systems by Takagi-Sugeno decomposition (Bezzaoucha, et al., 2013;Guerra, et al., 2015) is another possibility, as is the use of models using symmetries and semi-invariants (Menini & Tornambè, 2011), or the use of immersible techniques for systems transformed into a non-linear observer form (Back & Seo, 2008).
A large number of non-linear MISO systems with multiple inputs and a single output can be transformed into state equations using the form : with the following definitions : n : the order of the system of non-linear differential equations m : number of independant inputs u 1 ptq T : the vector ru 11 ptq, . . .u m1 ptqs of the m independent inputs yptq : the measurable output variable z T ptq : the state vector r z 1 ptq . . .z n ptq s d T : the vector of the output parameters of the system Φptq : the non-linear function of vector u 1 ptq of the inputs s i " zptq, u 1 ptq ‰ : one of the n non-linear functions of the state vector sptq.
Such systems are often found in nonlinear robotic systems in the form of trigonometric functions.Other systems contain non-linear polynomials (strange attractors, Bernouilli functions, non-linear springs), polynomial fractions, or various common simple functions . . .The n non-linear functions of vector sptq employ a vector of m independent inputs u 1 ptq, as well as the state vector zptq as input variables.Such a procedure allows amongst other possibilities the description of bi-linear systems.We limited ourselves in this study to continuous functions in all points of type C 1 .One considers that the measurable output is a linear combination of zptq, superimposed on a non-linear function Φ " u 1 ptq ‰ .For an engineer or physicist, many applications have such a form.Often, non-linear systems (1) are transformable in a regulation canonical form concieved by (Fliess, 1990), and are written : with the following definitions : u i ptq : the pi ´1q ´th temporal derivative of the vector u 1 ptq, either u i ptq T " r u 1i ptq, . . .u mi ptq s i " 2 . . .n Uptq " " u 1 ptq . . .u n ptq ‰ : the n ˆm input matrix, with the group of n vectors u i ptq x i ptq : pi ´1q th temporal derivative of x 1 ptq x T ptq : state vector r x 1 ptq, . . .x n ptq s c : the output parameters vector of the transformed system θ ď n : index of last coefficient c i ‰ 0 Ψ r xptq, Uptq s : a scalar non-linear C 1 function A : the n ˆn matrix of which the last line is zero.
Conversion of the transformed version (2) to the initial representation (1) is performed using : with gptq : the vector of n non-linear inverted transformation functions g i r xptq, Uptq s which link xptq to zptq.
In (Schwaller, Ensminger, Dresp-Langley, & Ragot, 2016) a new observer was proposed which was adapted to this transformed form, and which provided non-biased robust estimates of xptq.This is not always the case for estimates of zptq.Functions g i rxptqs (1b) permit linking xptq to zptq (2c) and are called inverted transformations.Because of the nonlinearity of gptq, small perturbations of estimates of xptq may be considerably increased and strongly disturb estimates of zptq.The main aim of this study was to solve this type of situation, by introducing the inverted observer transform functions g i rxptqs.Doing this, the resistance to observer noise is affected (Bodizs et al., 2011), and one obtains a tool capable of limiting its impact on estimates of zptq.
Definition 1 Let us define, for the moment, a normalised pulse ω o " 2π{T o , which introduces a new time scale τ for the representation of the transformed state of the system : and for the inverse transformation system : zpτq " gpτq (5a) with : τ " ω o t, 9 x n ptq " 9 x n pτq ω o n (6a) f pτq and gpτq are vectors with dimension n.In (4b), Φpτq " Φ " u 1 pτq ‰ " Φ " u 1 ptq ‰ .Equations ( 6) define time dilatation or retraction of the state representation and its new parameters, without changing the pattern of the signal x i pτq.For the function Ψ, this is translated by the relation of changing the following scale representation : The function r Ψ r xpτq, Upτq s is obtained by replacing every state or command variable by the corresponding one in (6) and dividing everything by ω o n .
Afterwards, the procedure can be separated into several steps: in section 2, the estimation of the state of the transformed system (4) is dealt with ; in section 3 a new observation method of the inverse transformation functions which permit estimation of state variables (1) is presented ; in section 4 this new approach is applied to observe a system of management of activated sludge in a purification station ; the study is concluded in section 5.
Structure of the Observer in Canonical Form
To begin with, let us isolate the componant x 1 pτq of (4b) which will subsequently serve to determine the observation error.
To obtain y 1 pτq, the estimation of variable x 1 pτq, three cases may be distinguished.For θ " 1 : For θ " 2, it becomes : In the most general case where θ ą 2, ypτq ´Φpτq is filtered by : To analyze the effect of the filter, we rewrite (4b) in scalar form, ignoring r c θ`1 . . .r c n , which are all zero : If ( 11) is inserted in (10a), ( 9) or ( 8) as a function of θ, it becomes : The Laplace transformation of (12a) gives the transfer function : To develop the rest, y 1 pτq is used to determine the observer error.
Definition 2 To generate state estimates vpτq for the system (4), a PI observer structure is defined in (Schwaller, Ensminger, Dresp-Langley, & Ragot, 2016) with : with q xpτq (14g) and p xpτq (14h) as two distinct state vectors of dimension n ´1, coupled using the matrices A (14n) and q A (14o) of dimension pn ´1q ˆpn ´1q.The vectors q h and p h are also of dimension n ´1.The matrix A is constructed using the Kronecker operator which puts the upper diagonal at 1.The parameters h i , i " 0 . . .n are the gains of the observer.
Figure 1 illustrates the functional diagram of such an observer of third order.
The augmented vector vpτq (14i),(14h) and ( 14g) is used as estimation of xpτq and as variable of the function r Ψ r vpτq, Upτq s (14f).The state p xpτq (14b) is an observer exploiting the observation error ∆y 1 pτq (14c) via the gains h i (14m) serving to correct the state distances between the system and its observer.
In figure 1, for example, we have : The choice of using two state variables p xpτq and q xpτq is motivated by the n ´1 successive integrations of 9 q x n pτq in which no p h ∆y 1 pτq re-injection error is involved.This allows an increase in the robustness of the estimations to the measurement noise, which in general affects the variable y 1 pτq.One thus overcomes a common weak point of high gain observations, i.e. their sensitivity to measurement noise.The second advantage comes from the non-linear function r Ψ rvpτq, Upτqs which is no longer subjected to the restrictive conditions used in (Schwaller, Ensminger, Dresp-Langley, & Ragot, 2013), and covers the ensemble of the systems described by (Fliess, 1990).The vector r f pτq (14d), of dimension n ´1, compensates the effects of f pτq, and of possible external exogenous disturbance of (2) using the integral component I 0 pτq (14e).One notes that at the second order, for a gain h 0 " 0 inhibiting the integrator I 0 , the observer becomes similar to that proposed by (Gauthier, Hammouri, & Othman, 1992) for a SISO system.
In (Schwaller, Ensminger, Dresp-Langley, & Ragot, 2016), a full analysis was performed in order to determine the dynamics of the observation error ∆y 1 pτq (14c) and its successive derivatives, to characterise stability conditions and also the exponential convergent nature of estimates vpτq.A mthod to synthesize parameters h 0 . . .h n was also proposed.
New observers definitions
In (5b), the inverted transform functions gpτq allow converting the system in the canonical form of regulation back to the original form (1). Using the estimates vpτq reconstructed by the observer ( 14), it is possible to define : (15) One thus obtains estimates p zptq of zptq (1).If the stability conditions (Theorem 1 of ( The n estimates p z i ptq can be used as reference inputs to observe n state variables q z i ptq which tend towards (15a).Their temporal derivatives tend towards 9 p zptq, which themselves tend towards 9 zptq (16).With the model ( 14), one defines n first order observers.Each is normalised by a pulse ω i which leads to its dimensionless time definition (17e), possesses its own Lipschitz constant, and its specific stability conditions that we have to find.Synthesising the gains h i (subsection 2.4 of (Schwaller, Ensminger, Dresp-Langley, & Ragot, 2016)) gives h 0 " 1 and h 1 " 2. The n observers are written : 9 q z i pτ i q " I i pτ i q `2 ∆z i pτ i q `q s i pτ i q (17a) 9 I i pτ i q " ∆z i pτ i q i " 1 . . .n (17b) ∆z i pτ i q " p z i pτ i q ´q z i pτ i q (17c) with q zptq " " q z 1 ptq . . .q z n ptq ‰ the vector of the estimations of p zptq ; q s i pτ i q is the normalised non-linear function of 9 q z i pτ i q. Figure 2 illustrates (15) and ( 17).The general calculation procedure is as follows : • estimation of vpτq (14i) after treatment of (14); • estimation of p zptq (15) ; • estimation of the n state distances (17c) ; • determination of the n non-linear functions q s i pτ i q (17d) to access the n terms 9 q z i pτ i q (17a) and 9 I i pτ i q (17b) ; • integration of the n equations (17a) to obtain q zptq.
The temporal derivative of (17c) and inserting (17a) in the rest obtained enables one to obtain the expression of ∆9 z i pτq : ∆9 z i pτ i q " 9 p z i pτ i q ´9 q z i pτ i q i " 1 . . .n (18a) " ∆ r Ψ i pτ i q ´Ii pτ i q ´2 ∆z i pτ i q (18b) ∆ r Ψ i pτ i q " p s i pτ i q ´q s i pτ i q (18c)
Dynamics of the Observer Errors
We now characterise the dynamics of the observer errors by searching the n differential equations of the distances ∆z i pτ i q.
Due to the presence of integrators I i pτ i q, an extra temporal derivative is necessary to obtain the differential equation of the distances ∆z i pτ i q.To do this, it is necessary to define the following augmented vectors : q z i pτ i q T " " q z i pτ i q 9 q z i pτ i q ı (19c) The temporal derivative of (18b) is written : and gives the scalar expression of the differential equations of the observation errors.Using notations (19) gives the matricial writing of (20a) in the form of state equations : Assuming that the non-linear functions 9 s i are at least locally Lipschitz in Zpτ i q, and uniformly bounded in Υpτ i q in an invariant set, they are associated with a Lipschitz constant L i : Applying the Lipschitz inequality to (20b) permits reduction to ∆Zpτ i q the number of useful variables to characterise the perturbing difference ∆ 9 r Ψ i pτ i q.For many systems, if functions 9 s i are not globally of a Lipschitz type, they can be locally or be transformed adequately into the Lipschitz type.
Convergence of State Observations
Now let us try to analyse the globally asymptotic development of the observation errors and to characterise the limiting stability conditions of each observer (17).
Theorem 1 Let us consider a MISO system decomposable as described in (4), for which the observer structures ( 14) and (17) are used, and related to each other by the inverted transform function (15a).If the system function 9 s i " p Zpτ i q, Υpτ i q ı is locally of the Lipschitz type in p Zpτ i q and uniformly bounded in Υpτ i q in an invariant set, with a Lipschitz constant L i (22), then the observer (17) will be locally stable if the Lipschitz constant L i satisfies the following conditions : If the system function 9 s i " p Zpτ i q, Υpτ i q ı is globally of the Lipschitz type, and if the Lipschitz constant L i satisfy (23), then the observers (17) will be globally asymptotically stable.
Proof.The proof of theorem 1 can be demonstrated by proving the stability of (21a) using an appropriate positive Lyapunov function, like the following quadratic function : v i pτ i q " ∆z i pτ i q T P i ∆z i pτ i q (24b) The P i lower triangular matrix are defined as positive and satisfying the Sylvester criteria, with (24c).The proof of convergence is linked to the study of the sign of the derivative of the candidate for a Lyapunov function.This is obtained after temporal derivation of (24a), and after placing (21a) in the result obtained for terms ∆9 z i pτ i q : An appropriate choice of ϕ i1 , ϕ i2 can provide negative diagonal coefficients for Q i .The criterion of semi-negativity of Sylvester is then respected, and the successive minors of Q i will be of opposite sign, ensuring the semi-negativity of the first member on the right of (25b).Verifying the sign of the second member on the right of (25b) involves increasing N i pτ i q using the inequalities of Schwartz and Lipschitz (22) : To determine the sign of 9 v i pτ i q function, one applies the following inequality : to (26c) to obtain the desired increase of N i pτ i q : In (28a) yields a positive lower triangular matrix R i (28b), the diagonal elements of which are written : With negative functions 9 v i pτ i q, adding together the diagonal terms of (25c) and (29), and imposing Q i `Ri ď 0, one obtains the conditions (23).The sum Q i `Ri yields an inferior triangular matrix that satisfies Sylvester criteria of seminegativity if inequalities (23a) and (23b) are satisfied.Then, if ∆ 9 r Ψ i pτ i q (20b) is Lipschitz (22), 9 v i pτ i q is semi-negative and (21a) is globally and asymptotically stable ; (21a) is locally stable if ( 22) is locally Lipschitz Using the (theorem 2, section 2.3, (Schwaller, Ensminger, Dresp-Langley, & Ragot, 2016)), it is easy to demonstrate that the observers (17) will be exponentially convergent.
Application to a Sludge Activation Model
Let us now illustrate the proposed observation method by applying it to a non-linear example with multiple inputs.
The variables z 1 ptq, z 2 ptq z 3 ptq represent the state of the reactor (figures 3(e),(f),(g)), respectively the concentration of rapidly biodegradable substrate, the concentration of dissolved oxygen, the particle concentration of biomass, with (34) its parameters, all known, and z 1 p0q " 4.1, z 2 p0q " 3.0, z 3 p0q " 867 the initial conditions.The sizes y 1 ptq, y 2 ptq (31b) represent the measurable outputs.As this application of the general procedure of transformation permits passage from systems (1) to (2) (Fliess, 1990) it will permit the use of an observer similar to that proposed in ( 14), associated with inverted transformation (15) and with observers (17).
Observation of the Inverted Transformation System
The inverted transformation system (43) serves to form the errors (14c) of three first order observers of the same type as those defined in (17), in order to estimate p zptq.
We now try to determine the Lipschitz constants that subsequently will allow defining the stability conditions of each observer.We thus start by looking for L in (36) using the same calculation method as that explained in ((Schwaller, Ensminger, Dresp-Langley, & Ragot, 2016), section 3.1).
Figure 3. Input variables and state variables of the bioreactor | 4,699.8 | 2017-03-05T00:00:00.000 | [
"Mathematics",
"Environmental Science",
"Engineering"
] |
Two matrix metalloproteinase inhibitors from scrophularia striata boiss.
Many species belonging to the Scrophularia genus have been used since ancient times as folk remedies for many medical conditions such as scrofulas, scabies, tumors, eczema, psoriasis, inflammations. The aim of this study was to characterize the matrix metalloproteinases (MMPs) inhibitor compounds of the Scrophularia striata extract by bio-guide fractionation. The aerial parts of S. striata were collected and different extracts were sequentially prepared with increasingly polar solvents. The MMPs inhibitory activity of the crude extract and its fractions were evaluated by the Zymoanalysis method. The pure compounds were purified from the active fraction by chromatography methods. Chemical structures were deduced by nuclear magnetic resonance and mass spectrometry. Two active compounds (acteoside and nepitrin) were identified by bio-guide fractionation. The inhibitory effects of nepitrin and acteoside at 20 µg/mL were about 56 and 18 percent, respectivly. The inhibitory effects of acteoside at 80 µg/mL were increased to about 73 percent. In summary, the results suggest that nepitrin effectively inhibited MMPs inhibitory activity at low concentrations, whereas acteoside showed inhibition at high concentrations.
Introduction
Metastasis, the major cause of cancer mortality, is a complex phenomenon in which tumor cells invade surrounding tissues, penetrate blood vessels and exit vessels at distant sites to form secondary tumors (1). Most investigators unanimously agree that matrix metalloproteinases (MMPs) are critical enzymes involved in many aspects of cancer, including tumor growth, invasion, metastasis (2-3-4) and neovascularization (5-6).
MMPs are a family of zinc-dependent endoproteinases that play pivotal roles in the dynamic remodeling of the extracelluar matrix. Based on substrate preference and structural homology, MMPs are sub-classified by their
Plant material
The aerial parts of S. striata were collected from plants growing in the northeastern part of Iran, in the Ruin region (1350 m above sea level) in May 2006 and were dried at room temperature. A sample was authenticated by Dr F. Attar, and a voucher specimen was preserved in the Faculty of Sciencesۥ Herbarium at Tehran University, Tehran, Iran (TUH no. 36501).
Extraction and isolation
Different extracts were sequentially prepared with increasingly polar solvents using 1 Kg of the dried and powdered aerial parts of the plant and 5 L of each solvent. The resulting extracts consisted of: Petroleum ether (6.6 g dry weight corresponding to 0.6%), chloroform (11.4 g dry weight corresponding to 1.1%), ethyl acetate (12.8 g dry weight corresponding to 1.2%) and 80% methanol extract (16.4 g dry weight corresponding to 1.6%).
The 80% methanol extract was subjected to polyamide column chromatography and eluted by water followed by water-methanol mixes of decreasing polarity. Six main fractions were obtained (a-f). TLC analysis was performed on silica gel using ethyl acetate-methanol-wateracetic acid in various proportions as the mobile phase. Compounds were visualized under UV light (254 and 365 nm) or by spraying the plates with anisaldehyde-sulfuric acid reagent (15). Fractions b and d significantly inhibited MMPs activity. Re-chromatography of the fractions b and d on Sephadex LH-20 (methanol), yielded 32 mg nepitrin and 28 mg acteoside as active compounds.
Cell culture
The wehi-164 cells were seeded in 96-well tissue culture plates. Cells were maintained in an RPMI-1640 medium supplemented with 5% fetal calf serum, plus antibiotics, at 5% CO 2 and saturated humidity.
Dose-response analysis
Two-fold dilution of plant extracts, column fractions and the reference drug (piroxicam) were added to triplicate samples of cultured cells. Untreated cells were used as controls. Cells were cultured overnight, and then subjected to a colorimetric assay. Cytotoxicity was expressed as functional groups into collagenases, gelatinases, stromelysins, matrilysins, membrane type MMPs (MT-MMPs) and other non-classified MMPs (7-8).
The genus Scrophularia, consisting of about 300 species, is one of the most important genera belonging to the Scrophulariaceae. Many species belonging to this genus have been used since ancient times as folk remedies for many medical conditions (scrofulas, scabies, tumors, eczema, psoriasis, inflammations, etc.) (9-10). For example, dried roots of S. ningpoensis Hemsl are used as antipyretic, febrifuge and antibacterial, as a remedy for evening fever, erythema, mouth dryness, constipation, prurigo, furunculosis, sore throat, ulcerous stomatitis, tonsillitis and in the treatment of cancer (11-12). S. deserti is used in traditional medicine as an antipyretic, a remedy for kidney diseases and tumors and lung cancer (13). Hajiaghaee et al. (2007) reported that a total extract of S. striata at 80 µg/mL concentration moderately inhibited growth of the wehi-164 cell line (37%), whereas at lower doses (down to 10 µg/mL) its cytotoxicity was negligible and the cell viability percentage was more than 70% (14).
In the present paper, the inhibitory effect of S. striata was studied using bio-guide fractionation. The active substances and their chemical structures were deduced by nuclear magnetic resonance (NMR) and mass spectrometry.
General
The 1 H and 13 C NMR spectra of the isolated compound were measured in CD 3 OD at 500 and 125 MHz, respectively, using a Bruker AC 500 Spectrophotometer (Germany). Mass spectra were taken on a Finnigan TSQ-Mat 70 (70 eV) Spectrometer.
the percentage of viable cells at different sample concentrations. IC 50 was calculated as the dose at which 50% of the cells died relative to the untreated cells. The corresponding supernatants of the cultured cells were used for zymoanalysis.
Colorimetric assay
In the cytotoxicity assay, cells in the exponential phase of growth phase were incubated for 24 h at 37 ºC with 5% CO 2 with a serial dilution of samples. Cell proliferation was evaluated by a modified crystal violet colorimetric assay (16). After each experiment, the cells were washed with ice-cold PBS and fixed in a 5% formaldehyde solution. Fixed cells were stained with 1% crystal violet and the stained cells were then layered and solubilized with a 33.3% acetic acid solution. The density of the purple product that developed was measured at 580 nm.
Zymoanalysis
This technique was used to detect gelatinase (collagenase type IV or matrix metalloproteinase type-2, MMP-2) and MMP-9 in conditioned media (17). Briefly, aliquots of conditioned media were subjected to electrophoresis in a gelatincontaining polyacrylamide gel, in the presence of sodium dodecyl sulfate (SDS) under nonreducing conditions. After electrophoresis, SDS was removed by repeated washing with Triton X-100. The gel slabs were incubated overnight at 37 ºC in a gelatinase-activating buffer and subsequently stained with Coomassie Brilliant Blue R 250 (Sigma, MA). After intensive destaining, proteolytic areas appeared as clear bands against a blue background. Using a gel documentation system, quantitative evaluations of both the surface and intensity of the lysis bands, on the basis of grey levels, were compared with the untreated control wells and expressed as a percentage of the «relative expression» of gelatinolytic activity. The IC 50 for the MMPs inhibitory effect was calculated as the doses at which 50% MMPs inhibition occurred relative to untreated control cells.
Statistical analysis
The differences in cell cytotoxicity and gelatinase zymography were compared using the Studentۥ s t-test. Values of p < 0.05 were considered significant.
Results and Discussion
In a previous study, we showed inhibitory effects of a total extract and sequential extracts of S. striata. An 80% methanol extract, compared with other extracts significantly inhibited MMPs activity and showed low cytotoxicity. The activity of different extracts and fractions are summarized in Table 1. The 80% methanol solution was subjected to activity-guided fractionation on a polyamide column and eluted with a watermethanol mixture. The active fractions were subjected to further purification on a Sephadex LH-20 column, with the result that two compounds with MMPs inhibitory activity were obtained (Figure 1).
The chemical structure confirmation of the components from the S. striata methanol extract was accomplished by comparing the obtained 1 H and 13 C NMR data to those previously published.
Compound 1 * p < 0.05. ** p < 0.01. *** p < 0.001. a Each data point was compared with control group and statistical analysis was performed with the Student's t-test.
The cytotoxicity and inhibitory effect of the active compounds of S. striata were evaluated in-vitro at four doses (10, 20, 40 and 80 µg/ mL) against the wehi-164 cell line. Both active compounds showed a direct dose-response as higher concentrations led to higher toxicity and inhibitory effects (Figure 2 and 3).
The invasion of wehi-164 cells was significantly inhibited at lower concentrations of nepitrin. The inhibitory effects of nepitrin at doses of 10 and 20 µg/mL were about 44 and 56 percent, respectively. Nepitrin at 2080-µg/ mL concentrations moderately inhibited MMPs activity, whereas at lower doses (down to 20 µg/ mL) its anti-invasive activity was substantial. Its cytotoxicity at lower doses was negligible and there was no significant difference between the control (0 µg/mL) and 10 µg/mL. The IC 50 values for cytotoxicity and zymoanalysis observed for the cell line are reported in Table 1.
The inhibitory effect of acteoside at lower doses (down to 20 µg/mL) was less than 18%, whereas at 40 and 80 µg/mL inhibition increased to about 47 and 73 percent, respectively. The viability percentage of acteoside at 10 µg/mL was about 81% and its cytotoxicity at 20, 40 and 80 µg/mL was negligible. The IC 50 value for zymoanalysis was calculated as 51.73 µg /mL.
In the present study, isolation of two MMP inhibitors was obtained using bioactivityguided fractionation of aerial parts of S. striata. Nepitrin effectively inhibited MMPs activity at low concentrations, whereas acteoside showed inhibition at high concentrations. Future work to extend these results should investigate the effect of these compounds on different cell lines, to gain a better perspective of their properties. | 2,203.6 | 2014-01-01T00:00:00.000 | [
"Biology"
] |
Computer-Aided Greenery Design—Prototype Green Structure Improving Human Health in Urban Ecosystem
Increasing population and urbanization, with climate change consequences, such as rising temperatures, influence public health and well-being. The search to improve the quality of life in cities becomes one of the priority objectives. A solution can be found in the role of greenery in an urban environment and its impact on human health. This opens a path toward experimentation on microclimate green structures that can be inserted into dense urban spaces providing human and environmental benefits. The article proposes an automated greenery design method combined with rapid prototyping for such interventions. A theoretical analysis of the problem preceded the introduction of the method. The research process was developed in accordance with the main objectives of the CDIO framework (Conceive, Design, Implement, and Operate) with the SiL (Software in the Loop) and HiL (Hardware in the Loop) methods. Moreover, the applied test model allows for complex evaluation in order to ensure quality and directions for further development.
Introduction
The urbanization and rise of the population are leading to climate change with all its consequences; for example, increased pollution and rising temperatures that affect human health conditions [1]. Therefore, the search to improve the quality of life becomes one of the emerging issues. One of the topics that gained recent scientific interest is the role of greenery in an urban environment and its impact on public health. Greening of the cities is discussed against the problems of air pollution and urban ventilation [2,3], cooling effects for buildings [4], improving their energy balance, and mitigating the heat island effects in cities by reduction of air and surface temperatures [5][6][7]. Urban vegetation enhances biodiversity and provides various ecosystem services [8,9]. Access to green spaces has proven to positively affect human health by relieving stress and improving the well-being of citizens [10][11][12]. Contact with nature supports our psychological and mental health and fosters social ties [13,14]. Along with personal and social benefits, human health contributes to economic growth and achieving sustainable development goals (SDG) [15].
These findings give the direction of research inquiries and architectural practices experimentation toward enhancing relationships with nature in cities and developing new typologies of designing with greenery such as ecological urbanism, biophilic urbanism, or landscape urbanism [16][17][18]. Urban forests and parks are considered key elements of the urban green infrastructure that deliver multiple health services; thus access to green spaces should inform urban policies and planning decisions [19]. A large group of studies relates to the methods of incorporating vegetation into building envelopes in the form of living walls, green facades, and green roofs [20,21]. Their impact on human health has been widely discussed-green walls may reduce air pollution and noise, mitigate urban heat island effects, and regulate humidity [22]. However, as some research studies reveal, the realization of such solutions could be time-consuming and often encounters While technological advancements facilitate the integration of new green fast fabricated prototypes into an urban structure, the idea of small temporary gardens in the city is not new. The inquiry into such forms of urban greenery intensified in the last decades with the emergence of environmental and health problems related to rapid urbanization processes. The environmental and landscape attributes, health benefits, and design challenges related to non-permanent landscapes have made them a creative and stimulating testing ground for researchers and designers. Raffaella Sini examines the historical evolution of the genre, exploring theories and strategies informing temporary small gardens. Key topics refer to temporary gardens as opportunities to work with live processes, and to gardening practice as a form of therapy and inclusion, which corresponds with social justice, public health, biodiversity, and ecology. Temporary gardens enrich the urban greenery forms and direct research studies toward a full range of new opportunities [34]. The recent findings on the health benefits of horticultural therapy, resulting not only from gardening but even viewing small gardens, particularly in the context of the aging society, reveal the need for working out innovative design solutions [35,36]. As Sini indicates, the time-limited design of small gardens affects the entire process of conceiving, building, experiencing, and managing green spaces; using short-term formats, anyone can invent, try, and experiment in a condensed experience of landscape [34].
In this context, the rapid prototyping methods and automatization of plant selection may bring new opportunities to the whole process. However, it is to be noted that automatization of the design process in the architectural field is a relatively new concept and only a few studies have taken up this topic [26]. Even though available software allows the use of visual algorithms or written scripts for the purpose of automatization processes, such methods are still not common in architecture and urban planning. Whereas, in the process of looking for the right answer to sustainability and health problems, in architecture, landscape architecture, and urban design, progress in pushing the boundaries is fundamentally associated with experimentation. Developing adequate research-by-design approaches is vital to provide architects, landscape architects, and urban designers with the knowledge and skills to meet the challenges of demographic, social, economic, environmental, and technological changes.
Thus, the purpose of this research is to propose a new, innovative method to develop easy-to-fabricate green structures that can be inserted into dense urban environments and thus contribute to public health and environmental benefits. The proposed approach is based on linking rapid prototyping with BIM tools and automated greenery selection. The presented study develops a new complex digital method of selection and design of greenery based on a parameter spreadsheet and environmental data. Parametrization of the chosen group of plants presented in this paper provides an initial step toward creating a database of plant species that can be used for algorithmic plant selection and distribution in further studies.
Methods
The research aimed to create a strategy for the development of a prototype microclimate installation based on green structure, rapid fabrication, and the method of automated greenery selection. The process of creating the test model for automated greenery design was based on elective research-by-design seminars organized for young researchers at the Faculty of Architecture at the Gdańsk University of Technology. The general process was conducted following the CDIO agenda and developed from its principles-SiL (software in the loop) and HiL (hardware in the loop) method developed for architectural research-by-design by Lucyna Nyka and Jan Cudzik [37].
The research was divided into three consecutive phases, starting with predesign, design, and ending with the fabrication and evaluation phase ( Figure 1). The strategy applied in the predesign research started with the digitalization of the site which was supported by environmental analysis of temperature range, soil type, moisture, and sun The research was divided into three consecutive phases, starting with predesig sign, and ending with the fabrication and evaluation phase ( Figure 1). The strateg plied in the predesign research started with the digitalization of the site which wa ported by environmental analysis of temperature range, soil type, moisture, and s posure. The analysis led to the determination of the scope of the design task and pa ters of plant species, materials, and technology. The basis for the automated greene lection process was the algorithm that used the RhinoCommon Library to develop r based on the combination of multiple parameters with the use of a visual editor, G hopper ® (Robert McNeel & Associates, Seattle, WA, USA). The digital analysis was supported by Ladybug ® , a plugin for Rhinoceros ® . Th cess led to necessary adjustments of automated greenery design (AGD) methods f purpose of the presented research study [26]. The AGD method takes into account p eters such as sun exposure, temperature range, moisture, maintenance, soil type, an reaction. The analysis of these parameters allowed for the automated selection of sp plants according to the location and characteristics of the created architectural form The design phase consists of four consecutive steps. The first one was the desig cess carried out by the involved researchers. The second was the revisions and sel of the design for further development. The selected design was then revised with a puter algorithm automatically creating greenery design scenarios that were adopte applied in the experiment. The last step of the design phase was the optimization of d fabrication. In the phase of fabrication and evaluation, the first step was the applicat The digital analysis was supported by Ladybug ® , a plugin for Rhinoceros ® . The process led to necessary adjustments of automated greenery design (AGD) methods for the purpose of the presented research study [26]. The AGD method takes into account parameters such as sun exposure, temperature range, moisture, maintenance, soil type, and soil reaction. The analysis of these parameters allowed for the automated selection of specific plants according to the location and characteristics of the created architectural form.
The design phase consists of four consecutive steps. The first one was the design process carried out by the involved researchers. The second was the revisions and selection of the design for further development. The selected design was then revised with a computer algorithm automatically creating greenery design scenarios that were adopted and applied in the experiment. The last step of the design phase was the optimization of digital fabrication. In the phase of fabrication and evaluation, the first step was the application of a digital production that was later followed by manual processing and securing the material with impregnation. This step allowed for on-site construction. In the last stage of the process, the design was evaluated.
Sample Scenario
The selection of location was based on general parameters: accessibility, good exposure to the sun, green surroundings, and safety. The chosen area is a part of the Gdańsk University of Technology campus, surrounded by sports facilities, the Pavilion of Architecture, and Steffens Park. The location is set between the railway line and 6 alley road ( Figure 2). a digital production that was later followed by manual processing and securing the material with impregnation. This step allowed for on-site construction. In the last stage of the process, the design was evaluated.
Sample Scenario
The selection of location was based on general parameters: accessibility, good exposure to the sun, green surroundings, and safety. The chosen area is a part of the Gdańsk University of Technology campus, surrounded by sports facilities, the Pavilion of Architecture, and Steffens Park. The location is set between the railway line and 6 alley road ( Figure 2). The digital environmental analysis was conducted with Rhinoceros ® and Grasshopper ® supported with the Ladybug ® plugin [38]. The source for environmental data was an EnergyPlus Weather file (.epw) located in the Nowy Port district that was within acceptable distance for such analysis [39]. This analysis consisted of sun hours throughout the year and a range of temperatures [40]. Moreover, on-site soil analysis with the humidity and soil reaction was conducted. Due to the plan of putting specific plants in pots, a suitable type of soil was selected according to plant needs.
Species Selection
For the purpose of the experiment, a group of climber plants was selected. These plants clothe walls and support foliage and flowers. Climbers cling by using tendrils, twining stems, stem roots, or sticky pads, while wall shrubs need to be tied to supports. Climbers mainly grow upwards and prefer sturdy supports to help them on their way [41]. This group of plants offers a range of sun-sensitivity from sun-loving to shade-loving, require relatively low maintenance, and are fast-growing species so that the effects of growth are noticeable even after one season [42]. Immersive spread of these plants invites planting participants to revisit the site of the experiment for further study. Because of these characteristics, climbers provide a group of plants that are well-fitted for a performance experiment in an architectural environment [43]. The digital environmental analysis was conducted with Rhinoceros ® and Grasshopper ® supported with the Ladybug ® plugin [38]. The source for environmental data was an EnergyPlus Weather file (.epw) located in the Nowy Port district that was within acceptable distance for such analysis [39]. This analysis consisted of sun hours throughout the year and a range of temperatures [40]. Moreover, on-site soil analysis with the humidity and soil reaction was conducted. Due to the plan of putting specific plants in pots, a suitable type of soil was selected according to plant needs.
Species Selection
For the purpose of the experiment, a group of climber plants was selected. These plants clothe walls and support foliage and flowers. Climbers cling by using tendrils, twining stems, stem roots, or sticky pads, while wall shrubs need to be tied to supports. Climbers mainly grow upwards and prefer sturdy supports to help them on their way [41]. This group of plants offers a range of sun-sensitivity from sun-loving to shade-loving, require relatively low maintenance, and are fast-growing species so that the effects of growth are noticeable even after one season [42]. Immersive spread of these plants invites planting participants to revisit the site of the experiment for further study. Because of these characteristics, climbers provide a group of plants that are well-fitted for a performance experiment in an architectural environment [43].
Many aspects of climbing plants have attracted scientific attention, from attachment mechanisms and general anatomy to genetic makeup and chemical properties [44]. Climbing vegetation growth has gained considerable research interest, from Darvin's studies [45] to modern-day experimentations [46]. This group of plants is a good fit when designing for better well-being in cities. They are applicable in regard to particular health and environmental benefits such as removing air pollution [47] or improving the cooling performance of buildings [48].
The research on planting preferences and characteristics of growth for particular species provided parameters that were used for the Automated Greenery Design method (AGD). The sources of the data were several research studies regarding vegetation [49][50][51][52][53][54]; in particular, the characteristics of plant growth were based on "The Plant Growth Planner" by Caroline Boisset [53] and preferred sun exposure was developed according to "Begrünte Architektur" by Rudi Baumann [54]. The final data is the outcome of the evaluation and selection of available plant parameters. The research aimed to analyze and provide the set of parameters that affect plant vegetation in relation to the architectural environment. The data used in the research was created by the Laboratory of Digital Technologies and Materials of the Future, Faculty of Architecture, Gdańsk University of Technology. Parameterization of plants' data conducted in the laboratory is part of a broader research program on the possibility of automation of design processes. An example of the gathered and curated data is presented in the table below (Table 1). 2D representation Green dot Green dot spline/point -
Photo
Many aspects of climbing plants have attracted scientific attention, from attachment mechanisms and general anatomy to genetic makeup and chemical properties [44]. Climbing vegetation growth has gained considerable research interest, from Darvin's studies [45] to modern-day experimentations [46]. This group of plants is a good fit when designing for better well-being in cities. They are applicable in regard to particular health and environmental benefits such as removing air pollution [47] or improving the cooling performance of buildings [48].
The research on planting preferences and characteristics of growth for particular species provided parameters that were used for the Automated Greenery Design method (AGD). The sources of the data were several research studies regarding vegetation [49][50][51][52][53][54]; in particular, the characteristics of plant growth were based on "The Plant Growth Planner" by Caroline Boisset [53] and preferred sun exposure was developed according to "Begrünte Architektur" by Rudi Baumann [54]. The final data is the outcome of the evaluation and selection of available plant parameters. The research aimed to analyze and provide the set of parameters that affect plant vegetation in relation to the architectural environment. The data used in the research was created by the Laboratory of Digital Technologies and Materials of the Future, Faculty of Architecture, Gdańsk University of Technology. Parameterization of plants' data conducted in the laboratory is part of a broader research program on the possibility of automation of design processes. An example of the gathered and curated data is presented in the table below (Table 1). Many aspects of climbing plants have attracted scientific attention, from attachment mechanisms and general anatomy to genetic makeup and chemical properties [44]. Climbing vegetation growth has gained considerable research interest, from Darvin's studies [45] to modern-day experimentations [46]. This group of plants is a good fit when designing for better well-being in cities. They are applicable in regard to particular health and environmental benefits such as removing air pollution [47] or improving the cooling performance of buildings [48].
The research on planting preferences and characteristics of growth for particular species provided parameters that were used for the Automated Greenery Design method (AGD). The sources of the data were several research studies regarding vegetation [49][50][51][52][53][54]; in particular, the characteristics of plant growth were based on "The Plant Growth Planner" by Caroline Boisset [53] and preferred sun exposure was developed according to "Begrünte Architektur" by Rudi Baumann [54]. The final data is the outcome of the evaluation and selection of available plant parameters. The research aimed to analyze and provide the set of parameters that affect plant vegetation in relation to the architectural environment. The data used in the research was created by the Laboratory of Digital Technologies and Materials of the Future, Faculty of Architecture, Gdańsk University of Technology. Parameterization of plants' data conducted in the laboratory is part of a broader research program on the possibility of automation of design processes. An example of the gathered and curated data is presented in the table below (Table 1). The group of plant species from the climbers family was chosen from an available database produced by the Laboratory of Digital Technologies and Materials of the Future. The most relevant criterion for this experiment was to provide a diversified spectrum of solar exposure so that any area on and around the planned installation can be matched with a particular species for planting. The chosen species are: Hedera helix
Parthenocissus
In subsequent steps of the study, plants' parameters in combination with climatic data allowed for the automatic analysis process to be carried out as an integrated element of spatial structure design.
Experimental Model Construction: Material and Fabrication Technology Selection
For the strategy of building and testing the prototype to work, adequate consideration of construction material was necessary. The main drivers behind the choice of materials for the construction of a prototype green structure were determined by the project's aim to implement the presence of greenery support in urban scenarios with minimum effort and maximum adjustment to the existing environment.
The other criteria for selecting the material were driven by level of complexity in terms of digital fabrication, accessibility, resilience to weather conditions, structural strength, ease of manual material processing, and low carbon footprint [55]. The considered materials were concrete, high-density Styrofoam, wood, plywood, carbon fiber, steel, and aluminum. After careful consideration plywood was selected. Plywood's structural strength makes it an excellent choice for constructing installations and CNC production which allows the designer almost an infinite array of options [56]. Plywood parts can be joined in various ways and the technological aspects of joinery can be easily learned on a basic level. The material is produced in many factories worldwide and therefore provides access worldwide [57]. The material should be adjusted to the production location according to the type of trees grown locally. The environmental parameters were highly important in the selection process in order to embed the potential for raising eco-awareness into the structure itself [58].
The selection of digital fabrication methods considered laser cutting, advanced robotics, 3D printing, and CNC. The complexity of the use of robotics, limitations of scale in 3D printing, and impossibility to cut thick structural parts in wood with the use of laser cutting, all excluded these options for the production method. On the other hand, accessibility, ease of use, ability to cut structural parts in plywood, and low-energy post-production manual process all established the use of Computer Numerical Control machines as an option of choice [59]. CNC production allows plywood parts to be cut and routed with extreme accuracy, opening up a whole range of design possibilities for the design process. CNC cutting is currently being utilized by architects to make a whole range of different objects ranging from furniture, kitchens, built-in joinery, and prefabricated housing [60].
There are many different types of plywood available on today's market, ranging from sheets formed in slow-grown dense tropical hardwoods to fast-grown lightweight softwoods. Since the installation was intended for an outdoor site, due to its ability to balance temperature and humidity variations as well as provide sufficient constructional properties, we have selected water-resistant birch plywood 18 mm thick and additional impregnation with environmentally neutral products.
Design Process
The strategy of integrating young researchers into research and development tasks provided numerous solutions for the intended realization of the microclimate structure. (Figure 3). Priority of project selection was outlined in the design brief and included the project's qualitative merits including plant distribution preferences, constructional innovation, buildability, and benefits to a surrounding environment. Instructions also called for a project that should demonstrate consideration of aesthetic, technical, functional, economic, ecological, and sustainability requirements to be encompassed in the design. The most successful design proposition was selected for further development and fabrication phase.
Architectural design is a process that integrates several different disciplines though some of them are not well represented in early-stages including greenery design. That often becomes only a part of the final design stages whereas it should be a part of all stages and therefore affect the decision-making process. The design agenda aimed mainly at solving this issue. The core idea was to allow for experimentation in search of the most versatile structures providing diversified conditions for selected species of plants. The test group consisted of 15 young researchers that were divided into six teams. Each team had the same goal, which was to design a form that would provide support and conditions for four types of climbing plants with different needs. Each group started with sketches which got reviewed and then worked on both digital and physical models which allowed for a better understanding of the form and better evaluation of the design process according to CDIO principia. The final presentation of proposals led to a discussion about achieved results and possible changes.
project's qualitative merits including plant distribution preferences, constructional innovation, buildability, and benefits to a surrounding environment. Instructions also called for a project that should demonstrate consideration of aesthetic, technical, functional, economic, ecological, and sustainability requirements to be encompassed in the design. The most successful design proposition was selected for further development and fabrication phase. Architectural design is a process that integrates several different disciplines though some of them are not well represented in early-stages including greenery design. That often becomes only a part of the final design stages whereas it should be a part of all stages and therefore affect the decision-making process. The design agenda aimed mainly at solving this issue. The core idea was to allow for experimentation in search of the most versatile structures providing diversified conditions for selected species of plants. The test group consisted of 15 young researchers that were divided into six teams. Each team had the same goal, which was to design a form that would provide support and conditions for four types of climbing plants with different needs. Each group started with sketches which got reviewed and then worked on both digital and physical models which allowed for a better understanding of the form and better evaluation of the design process according to CDIO principia. The final presentation of proposals led to a discussion about achieved results and possible changes.
All of the projects provided acceptable solutions to design problems. The concept selected for the construction, however, distinguished itself from the others with the quality of achieved design goals. Firstly, the chosen proposal presented the most diverse distribution of greenery positions on the structure relative to different levels of sun exposure. All of the projects provided acceptable solutions to design problems. The concept selected for the construction, however, distinguished itself from the others with the quality of achieved design goals. Firstly, the chosen proposal presented the most diverse distribution of greenery positions on the structure relative to different levels of sun exposure. This allowed for the automatic plant selection algorithm to use every plant species from the provided list when matching the designated spots with different kinds of vegetation. Secondly, the chosen design, when compared to the other proposals, offered the most efficient material usage. It used only 5 sheets of plywood, whereas most of the other designs used exceedingly more. Thirdly, the selected structure was the easiest to build. It used the smallest number of timber joints out of all proposals and also, in contrast with the other designs, provided specifically designed holes for the pots, therefore ensuring the easiest construction and fast greenery planting.
The Automated Selection Process
The automated greenery design (AGD) is a method for including plants in early-stage design processes, developed at the Digital Technologies and Materials of the Future Laboratory of the Gdańsk University of Technology. The AGD (Figure 4) allows for automated plant selection based on plant species parameters combined with site environmental analysis. Parameters that are considered can be adjusted according to the designers' needs and characteristics of the project. In the presented experimental model, the following parameters were taken into consideration: sun exposure, temperature range, moisture, maintenance, soil type, and soil reaction. boratory of the Gdańsk University of Technology. The AGD (Figure 4) allows mated plant selection based on plant species parameters combined with site env tal analysis. Parameters that are considered can be adjusted according to the d needs and characteristics of the project. In the presented experimental model, th ing parameters were taken into consideration: sun exposure, temperature range, maintenance, soil type, and soil reaction. To provide the parameters for the algorithm, a database was developed. base has two forms of representation. The first one consisted of a series of para the form of a spreadsheet in .csv file format that was applied to the algorithm. T one was developed in the form of a table for presenting the parameters to the and user ( Table 1). The environmental evaluation was based on on-site and dig sis. The soil type and reaction were measured at the selected location. The humid temperature range, and sun exposure ( Figure 5) were calculated using digital an environmental data from the EnergyPlus Weather file (.epw) located in the N district. To provide the parameters for the algorithm, a database was developed. The database has two forms of representation. The first one consisted of a series of parameters in the form of a spreadsheet in .csv file format that was applied to the algorithm. The second one was developed in the form of a table for presenting the parameters to the designer and user ( Table 1). The environmental evaluation was based on on-site and digital analysis. The soil type and reaction were measured at the selected location. The humidity range, temperature range, and sun exposure ( Figure 5) were calculated using digital analysis on environmental data from the EnergyPlus Weather file (.epw) located in the Nowy Port district. The automation was conducted with Rhinoceros ® , Grasshopper ® , and Lady tools. The outcome of the process was the optimal scenario for plant selection comp for the chosen design of the structure (Figure 6). In the computational process, the The automation was conducted with Rhinoceros ® , Grasshopper ® , and Ladybug ® tools. The outcome of the process was the optimal scenario for plant selection computed for the chosen design of the structure (Figure 6). In the computational process, the environmental analysis was combined with plant parameters that allowed for the solution to be presented in the form of a schematic view that points out a specific plant location. Subsequently, this scenario was applied in a prototype experimental model. The automation was conducted with Rhinoceros ® , Grasshopper ® , and Ladybug ® tools. The outcome of the process was the optimal scenario for plant selection computed for the chosen design of the structure (Figure 6). In the computational process, the environmental analysis was combined with plant parameters that allowed for the solution to be presented in the form of a schematic view that points out a specific plant location. Subsequently, this scenario was applied in a prototype experimental model.
Fabrication
After selecting the final form, a 3D model of the microclimate structure was recreated with the use of the McNeel Rhinoceros ® software (Version 6; Robert McNeel & Associates, Seattle, WA, USA) according to digital fabrication needs. The evaluation was then carried out to ensure buildability and optimization. The 3D model was revised to accommodate tolerances into joints in plywood and an additional cutting path was added to exclude the rounding of the tool bit out of the join slots (Figure 7) [61].
The next stage consisted of flattening parts of the installation and nesting these elements onto 1525 × 1525 mm plywood sheets to minimize material waste (Figure 8) [62]. All the parts were numbered and placed on sheets according to the building schedule to allow construction to start while other parts were being cut on the CNC plotter, therefore minimizing the time of construction. The aim of such an approach was to examine the possible solution for further development of model fabrication management. out to ensure buildability and optimization. The 3D model was revised to accommod tolerances into joints in plywood and an additional cutting path was added to exclude rounding of the tool bit out of the join slots (Figure 7) [61]. When all the boards of plywood were prepared for cutting, the CNC movements had to be planned in the form of a g-code. The tool path was generated using RhinoCam ® software (Version 1.0; MecSoft Corporation, Dana Point, CA, USA), which is a plugin for the Rhinoceros ® (Figure 9). G-code was exported from the software in the form of a text file, consisting of a list of coordinates, and then imported into the CNC plotter [63]. The tool bit of choice was an up-down drill, 8 mm in diameter, which is used for fast and deep cutting action with minimal post-production needs.
ic Health 2023, 20, x FOR PEER REVIEW 13 of 21 Figure 9. Tool path generation.
While CNC milling was in progress the manual processing of already-cut parts began ( Figure 10). The elements of the structure had to be sanded and then a coat of varnish was applied for additional water protection. As soon as the parts dried, they were joined with the rest of the prepared elements ( Figure 11). While CNC milling was in progress the manual processing of already-cut parts began ( Figure 10). The elements of the structure had to be sanded and then a coat of varnish was applied for additional water protection. As soon as the parts dried, they were joined with the rest of the prepared elements ( Figure 11). While CNC milling was in progress the manual processing of already-cut par ( Figure 10). The elements of the structure had to be sanded and then a coat of varn applied for additional water protection. As soon as the parts dried, they were join the rest of the prepared elements ( Figure 11). The process continued until the assembly was completed. The process of fab was completed within one working day according to schedule, without any prio ration, and without occurring errors. Construction was assembled on-site ( The process continued until the assembly was completed. The process of fabrication was completed within one working day according to schedule, without any prior preparation, and without occurring errors. Construction was assembled on-site ( Figure 12). Moreover, throughout the time of process, all vegetation was planted in pots according to design decisions and to automatic plant selection protocol. The pots were filled with a layer of gravelite mixed with sand and then with a layer of soil. The whole mix was enriched with a composition of ecological fertilizers containing nitrogen and phosphorus compounds. The chosen climbers were planted in the pots and their elements were tied to the structure to ensure proper support.
The process continued until the assembly was completed. The process of was completed within one working day according to schedule, without any p ration, and without occurring errors. Construction was assembled on-site Moreover, throughout the time of process, all vegetation was planted in pots a design decisions and to automatic plant selection protocol. The pots were f layer of gravelite mixed with sand and then with a layer of soil. The whole m riched with a composition of ecological fertilizers containing nitrogen and p compounds. The chosen climbers were planted in the pots and their elemen to the structure to ensure proper support.
Final Installation-Maintenance, Inspections, and Verification of the Method
The created prototype is an effect of analysis and searches for the most sign solution for the green structure that can be built in a relatively short tim was not to automate the entire form-finding process but to connect the traditi
Final Installation-Maintenance, Inspections, and Verification of the Method
The created prototype is an effect of analysis and searches for the most suitable design solution for the green structure that can be built in a relatively short time. The aim was not to automate the entire form-finding process but to connect the traditional design approach (manual design or BIM design) with the automated greenery design system. This allows for evaluating the effectiveness of the automated vegetation selection method and testing it on a controlled laboratory scale. The outcome is promising and shows the applicability and scalability of the method for more complex projects.
The design process of the created structure was set according to the CDIO model [37] with the application of SiL-and HiL-verification loops. The inspections allowed for a better understanding of growing patterns and possible risks that accompany them. The experimental prototype becomes evidence of the durability of selected materials and plants. Regarding the possibility of creating the installation, it was assumed that it should not exceed a working week. The design process was set accordingly to enable digital fabrication. The design was aiming for the maximum reduction of unnecessary connecting elements and waste. The above criteria are assessed positively, the form was designed, analyzed, and produced within 5 days, and the entire fabrication was carried out using a CNC machine. Manual processing was only a minor percentage of the whole task and could be reduced in the future.
As testing through the seasons (spring, summer, and fall) showed, the planted vegetation has grown according to the described growth pattern, and the conditions allowed for the diversification of plants in terms of light exposure ( Figure 13). The placement of particular plants was set correctly according to the environmental needs with all plants completing the full vegetation circle. The prototype needed additional maintenance during the unexpectedly dry season which proves the need for site inspection. Application of the method to the creation process of the experimental prototype made it possible to verify the effectiveness and efficiency of the vegetation selection method at the early stages of designing and to verify the parameterization of features. An important element of the supervision and observation process is also the possibility to draw conclusions regarding any additional parameters that should be considered in the process of parameterization of greenery features for the purposes of automation of design processes.
Discussion
According to the presented research method, temporary or seasonal installations can be designed and later developed into larger and more complex spatial systems. A prototype-based process carried out in this way allows for quick and precise testing and customization. The presented method is characterized by the ability to quickly respond to emerging problems in the city, such as heat islands, and the inability to introduce permanent green urban tissue. The main aspect of this solution is the possibility of the application of a digital selection of plants at the early design stage. Due to the implementation of tools that allow for automatic design, the proposed method may be applied by local stakeholders. It does not require specialist knowledge in the field of botany and computational design. This would, however, require creating a library of plant parameters for particular geographical locations according to specific plants' functionalities as an easily accessible open-source database. Developed for the purpose of this study, the research framework library can be freely enlarged, both in terms of additional parameters and other plant species.
The method of plant selection also allows for easy integration with digital methods of rapid prototyping and the BIM-based design process. This would allow for modeling and evaluating the chosen data through analysis. For instance, the thermal radiation of public spaces could determine where the easy-to-fabricate microclimate prototype structures should be located [64]. Research on the applicability of the proposed solution to reduce the negative effects of urban heat islands is of utmost importance given that, according to the European Climate and Health observatory, heat-related mortality has almost doubled in the last 20 years [65,66].
The automated selection process may be focused in future research on the potential of vegetation to accumulate water or improve soil quality by the absorption of heavy metals [67]. For further investigations, the new parameters could be added to the plants' cards. This could include plants' ability to deposit and absorb particulate matter [68]. Since mitigating urban heat islands depends on particular plant qualities, such as Leaf Area Index (LAI), this property could also be included in the library of parameters [20]. The proposed method is easy to apply in building interiors, where the main indicators for health and comfort include indoor air quality as well as thermal and humid conditions [69].
The great advantage of the method is the ability to create independent structures that can be easily located in problematic places and then moved, changed, or expanded. The method allows for the appropriate selection of plant species for a particular site. It also allows for verification in terms of selected parameters, depending on the purpose of implementation. For instance, in particularly sensitive locations, such as healthcare or daycare centers, the plants could be selected based on their unique health-promoting properties such as their ability to reduce anxiety [70]. As Muahram et. al. proves, even exposition to various spectrums of green colors mitigates stress [22] but closer insights into molecular-scale modeling reveal much more; for instance, how human health may benefit from the therapeutic properties of particular types of vegetation on the cellular level [16]. The proposed microclimate structure can be adapted to different types of designs and functions, such as public transport stops or shading canopies. The solution may also include a number of other requirements, such as maintenance demands or limitations.
Conclusions
The proposition of the method developed for the small-scale green structure prototype provides an important insight into the challenges that go along with the incorporation of vegetation into the design process. In the era of climate change characterized by rising temperatures, expansions of cities, shrinking areas of urban greenery, and demographic changes such as the aging population, the experimental prototype can become a possible rapid solution that could increase the quality of life in cities and bring benefits to human health. The creation of the experimental prototype made it possible to verify the effectiveness and efficiency of the vegetation selection method at the early stages of the project. Moreover, it is highly important to raise the designers' awareness of the need and opportunities concerning the incorporation of greenery design in the early stages of the design processes, not only for the small-scale prototypes but also buildings and public spaces.
Additionally, the proposed computer method would allow for a more precise and knowledge-based holistic design of greenery. The achieved solution would require consultation with a landscape architect, gardener, or horticulture therapist. However, it would constitute a good basis, supported by hard data, for the development of the concept at later stages of design. The later stages may focus on the search for additional species for enhancing biodiversity in flora and fauna, stimulating olfactory sensations, or increasing effectiveness for air pollutant reduction, depending on the needs. The proposed method could become a response to the often over-complicated interface or redundant data provided at various stages of the work. The plant cards created for the system could consist of selected greenery parameters (SGP), considered during computer analyzes, but could also include an additional full greenery parameters table (FGP) allowing for additional verification of special cases. In the authors' opinion, creating cards and their variants is the greatest challenge, and an attempt to parametrize individual requirements may be a great research task. However, the advances in this field of research could contribute to creating more healthy urban environments and therefore benefit public health. | 9,574.6 | 2023-01-01T00:00:00.000 | [
"Materials Science"
] |
Dintor: functional annotation of genomic and proteomic data
Background During the last decade, a great number of extremely valuable large-scale genomics and proteomics datasets have become available to the research community. In addition, dropping costs for conducting high-throughput sequencing experiments and the option to outsource them considerably contribute to an increasing number of researchers becoming active in this field. Even though various computational approaches have been developed to analyze these data, it is still a laborious task involving prudent integration of many heterogeneous and frequently updated data sources, creating a barrier for interested scientists to accomplish their own analysis. Results We have implemented Dintor, a data integration framework that provides a set of over 30 tools to assist researchers in the exploration of genomics and proteomics datasets. Each of the tools solves a particular task and several tools can be combined into data processing pipelines. Dintor covers a wide range of frequently required functionalities, from gene identifier conversions and orthology mappings to functional annotation of proteins and genetic variants up to candidate gene prioritization and Gene Ontology-based gene set enrichment analysis. Since the tools operate on constantly changing datasets, we provide a mechanism to unambiguously link tools with different versions of archived datasets, which guarantees reproducible results for future tool invocations. We demonstrate a selection of Dintor’s capabilities by analyzing datasets from four representative publications. The open source software can be downloaded and installed on a local Unix machine. For reasons of data privacy it can be configured to retrieve local data only. In addition, the Dintor tools are available on our public Galaxy web service at http://dintor.eurac.edu. Conclusions Dintor is a computational annotation framework for the analysis of genomic and proteomic datasets, providing a rich set of tools that cover the most frequently encountered tasks. A major advantage is its capability to consistently handle multiple versions of tool-associated datasets, supporting the researcher in delivering reproducible results. Electronic supplementary material The online version of this article (doi:10.1186/s12864-015-2279-5) contains supplementary material, which is available to authorized users.
, [4], [5] Gcoords2gcoords Convenience script to convert between Dintor genomic coordinates (GC) and other commonly used GC formats. gcoords2genes Query for human genes in the vicinity of a position on the genome, usually addressing a variation. Output includes the Ensembl human gene ID and the distance to the gene, its strand and a distancebased rank. gcoords2ld Compute linkage disequilibrium (LD) between pairs of GCs or a GC and a gene. Outputs D' and r 2 measures. For a GC/gene pair, all SNPs from the gene are taken and the maximum LD measure is reported. Calculation can be restricted to certain populations available in the 1000 Genomes/HapMap projects. [6] gcoords2reg Query for regulatory regions from the Encode project for a position on the genome. The level of output detail can be chosen and reflects the way Encode data are organized. [7] gcoords2snp Inverse function of snp2gcoords. Queries a position on the genome for dbSNP entries and outputs any rs* IDs found. [8] gcoordsconservation Query if a given position on the genome is located in a conserved regions across a selectable set of organisms. Depending on the choice of conservation assignment, output is either binary (conserved or not) or a GERP score.
[9], [10] gene2canonexons Retrieve a list of canonical exons for a human gene. Useful for producing input for Illumina DesignStudio software. liftgcoords Lift genomic coordinates from a previous human genome release to the current.
snp2gcoords
Convert dbSNP rs* IDs to genomic coordinates as used in Dintor. Output includes Ensembl quality control flags, reference and alternate alleles, strand information, and evidence backing up the SNP. Inverse functionality is given by gcoords2snp. [8] tbl2tbl Import filter for tabular data. Used to import arbitrarily formatted text-based tables into Dintor's tabseparated format. tblsubmerge Merge a table based on a column's unique identifier. Rows with the same entry in a specified column will be joined into a single row. tblsubsplit Split single or multiple cells into separate rows, optionally adding an index column that can be used to undo this operation by calling tblsubmerge. HSEnsgProteinMapper Human gene ↔︎ protein mapping tool. Maps between any combination of Ensembl gene ID, consensus coding sequence (CCDS), Ensembl transcript ID, Ensembl protein ID, UniProt SwissProt or Trembl accession number or entry name. A common use case is mapping from Ensembl gene IDs (or transcript IDs) derived from Dintor tools to UniProt or CCDS.
[13], [14] HSGeneOrthologyMapper Derive orthology information between human genes and the most common scientific model organisms, fruit fly, mouse, and worm. Orthology mappings are based on Ensembl. [15]
Interval2Genes
List human genes contained in a genomic region specified by a pair of begin/end GCs. Additionally, it can be used to relate a genomic position (as originating from a variation) to the genes contained in an LD block output by the Pos2LDBlock tool.
Pos2LDBlock
Assign LD-based haplotype blocks to a position on the genome. The output encodes the relationship between the query position and the LD-haplotype block. [16] TableJoiner Join two tables on a common column. Unlike the Unix join command, this tool works on arbitrary, unsorted, tab-separated tables and allows transferring subsets of columns from the joined table.
VCF2Dint
Import filter for variant call format (VCF) files into Dintor tables.
Dintor tool name Description References ClinVarAnnotator
Output hits in NCBI ClinVar database for a GC, a GC with reference and alternate allele, or an interval on a chromosome (e.g. derived from Interval2Genes). [17] DrugBankAnnotator Retrieve DrugBank information for UniProt accession numbers. Lists drugs associated with proteins identified by their respective UniProt accession numbers. [18] GOAnnotator Query Gene Ontology (GO) for either GO terms and their descendants or for GO terms associated with UniProt accession numbers. A variety of information can be retrieved, such as GO term names, evidence codes, and ontology name. Filters exist to limit the number of terms to a certain depth in the GO graph, to ontologies, to certain types of edges, and to high quality SwissProt entries. [19] HGMDAnnotator [not available on public Galaxy web server] The human gene mutation database (HGMD) contains manual annotations of human gene mutations. Due to licensing restrictions, access to this database is only available as a command line tool interfacing a purchased HGMD MySQL database. The tool itself has an interface comparable to ClinVarAnnotator, and a Galaxy interface is ready for license holders running their own Galaxy server. [20]
HSGeneAtlas
Retrieve tissue-specific gene expression for human genes using the Genomics Institute of the Novartis Research Foundation (GNF) Gene Atlas. Filters are available for gene over-, and/or under-expression and tissue types. [21] InteractionAnnotator Find protein interaction partners using the iRefIndex database. Additional data characterizing the interactions, such as external references and experimental detection techniques, can optionally be output. Interactions may be restricted to a panel of predefined genes/proteins. [22] PharmaADMEntor Highlight mutations in an industry-initiated database of genetic biomarkers reliably involved in drug metabolism. [23] ReactomeAnnotator Retrieve information about pathways, reactions, and participating molecules from the Reactome database, taking into account the hierarchical (parent-child) structure of the data. The tool can be queried by UniProt accession numbers, Reactome identifiers, or free text; the output may be restricted to a predefined panel. [24]
Dintor tool name Description References GOEnricher
Perform GO term-based gene set enrichment analysis. Enrichment can be performed on any of the three ontologies (biological process, molecular function, and cellular component). Correction for multiple hypothesis testing and result clustering are available. Enriched GO terms are usually based on the set originating from all genes, but can also be broken down to genes.
[25], [26] GOFunSim Compute pairwise protein functional similarity. The tool offers calculation of five different functional similarity measures based on six different semantic similarity measures. In addition, functional similarity can be computed between a list of proteins and predefined panel of (usually related) proteins. Furthermore, semantic similarity can be derived for pairs of GO terms. Graph-based GO term information content can also be output. [27] MendelianFilter Remove variants that do not comply with a certain mode of Mendelian inheritance. The tool operates on a multi-sample VCF file and furthermore requires relatedness be provided in a pedigree (PED format). Filtering is possible for autosomal dominant or recessive, X-linked dominant or recessive, and mitochondrial linked inheritance.
[28], [29] MetaRanker Given an object (e.g. gene) associated with multiple scores, each one in a single column of a table, compute a single, rank-based score from these columns. This module is used in the final ranking provided by Prioritizer. Columns may contain missing values, the ordering of a column's content can be specified individually, and the final rank calculation allows weighting the contributing columns. [30] Prioritizer Performs candidate gene prioritization by a guilt-by-association approach. Candidate genes are compared to a user-defined panel of related genes (e.g. disease associated) by the following Dintor tools: • InteractionAnnotator: Does the candidate gene interact with a panel gene?
• ReactomeAnnotator: Does the candidate gene share pathways with the panel genes?
• GOFunSim: Is there high functional similarity between the candidate gene and genes from the panel?
• GOAnnotator: Is the candidate gene involved in similar GO classes as the panel genes? [31] | 2,122.4 | 2015-12-01T00:00:00.000 | [
"Biology",
"Computer Science"
] |
Numerical Simulation of the Dynamics of Listeria Monocytogenes Biofilms †
A biofilm is a layer of microorganisms attached to a surface and protected by a matrix of exopolysaccharides. Biofilm structures difficult the removal of microorganisms, thus the study of the type of structures formed throughout a biofilm life cycle is key to design elimination techniques. Also, the study of the inner mechanisms of a biofilm system is of the utmost importance in order to prevent harmful biofilms formation and enhance the properties of beneficial biofilms. This study must be achieved through the combination of mathematical modelling and experimental studies. Our work focuses on the study of biofilms formed by Listeria monocytogenes, a pathogen bacteria, specially relevant in food industry. Listeria is highly resistant to biocides and appears in common food surfaces even after decontamination processes. Their biofilms can develop quite different structures, from flat biofilms to clustered or honeycomb structures. In the present work, we develop 1D and 2D models that simulate the dynamics of biofilms formed by different strains of L. monocytogenes. All this models are solved with efficient numerical methods and robust numerical techniques, such as the Level Set method. The numerical re sults are compared with the experimental measurements obtained in the Instituto de Investigaciones Marinas, CSIC (Vigo, Spain), and the Micalis Institute, INRA (Massy, France).
Introduction
Listeria monocytogenes is a pathogenic bacteria responsible for outbreaks of listeriosis.The main mode of transmission of this pathogen to humans is the consumption of contaminated food through contact with unhygienic work surfaces and facilities where L. monocytogenes can form biofilms [1].
Biofilms structure determines the main physiological processes related to persistence and resistance.Therefore, structure characterization is critical to design cost effective and environmentally friendly disinfection techniques [2].Confocal laser scanner microscopy (CLSM) allows for in vivo and in situ biofilms observation.
In parallel to the experimental studies, the use of efficient mathematical models allows the prediction of the biofilm evolution for particular values of the involved parameters associated to different conditions.Having in view the experimental dynamics of the particular biofilm formed by the L1A1 L. monocytogenes strain, we start by considering the most successful 1D continuum model studied in the recent work [2].With the knowledge acquired in the 1D model, a 2D continuum multispecies model is developed [3] so that we are able to describe several dynamics shown by different L. monocytogenes strains.Both models are solved numerically by applying efficient numerical techniques such as Crank-Nicolson schemes, WENO methods or the Level Set method [3,4].The numerical results that arise are compared with the experimental measurements obtained in the IIM-CSIC (Vigo, Spain), and the Micalis Institute, INRA (Massy, France).
One-Dimensional Model
To elucidate the mechanisms explaining the life cycle of the biofilms formed by L1A1 strain we analysed several models until reaching the most successful one [2].Unknown parameters from the model were estimated using data fitting techniques within the AMIGO2 toolbox [5].The model is a 1D deterministic reaction-diffusion model.It consist of a set of (non-linear) partial differential equations (PDEs) which describe the spatio-temporal dynamics of biomass and nutrients.The key features of the model are:
•
There is a sharp front of biomass at the bulk/solid transition.
•
Biomass density can not exceed a maximum bound which is a parameter of the model.
•
Biomass production is due to nutrient consumption.
•
Nutrients diffuse in the bulk and in the biofilm with different diffusion constants.
•
The detachment is related to biofilm ageing.
All in all, the model is described by the following equations: completed with appropriate initial and boundary conditions.Equation (1) describes the nutrients dynamics whereas Equation (2) describes the biomass dynamics.
Two-Dimensional Model
With the insights provided by the 1D case, a two-dimensional model is built so that it is able to describe the dynamics of the L1A1 strain as well as the clustered or honeycomb patterns presented by other strains such as the CECT 5873 [6].The proposed model is a deterministic multi-species model of the W-G type.The key hypotheses are:
•
Biofilm described as a viscous fluid.
•
Nutrients and biomasses concentrations governed by a mass conservation law.
•
Active and inactive biomasses are of the same microbial species and incompressible.
•
The time scale for the biomass-related processes is much slower that for the nutrients-related.
•
Nutrients are diluted in the media.Biomass exists only inside the biofilm.
•
The detachment is related to cells death and the degradation of the extracellular DNA, i.e., to biofilm ageing.
All in all, the model is described by the following equations: plus appropriate initial and boundary conditions.Equation ( 3) describes the nutrients dynamics and Equation ( 4) describes the biofilm expansion growth pressure.Equations ( 5) and ( 6) are related to the level set method.Equations ( 7) and ( 8) are the active and inactive biomass dynamics equations.
Results
After solving both models numerically with the appropriate optimal model parameters, the results yielded are presented in Figures 1 and 2. Starting with the 1D case, Figure 1 it can be noted that biomass thickness is slowed down reaching its peak around 100 h.Nutrients are consumed until the nutrients impairing mechanism starts, preventing biomass from consuming all the nutrients in the domain.Also, it can be observed how the massive detachment happens in the final stage.Results reveal that the model is in clear agreement with the experimental data.Therefore concluding that the life cycle of L1A1 L. monocytogenes under the tested experimental conditions may be explained by taking into account impaired nutrients uptake and a massive detachment due to biofilm ageing.
As for the 2D model, Figure 2 shows the different dynamics achieved by the model with appropriate modifications for the model parameters.The results for the flat biofilms correspond to the dynamics of L1A1 L. monocytogenes and are in good agreement with the experimental measurements.On the other hand, the results for the clustered biofilms and honeycomb patterns represent the dynamics of the CECT 5873 L. monocytogenes strain at different stages of its life cycle, as can be seen in Figure 2, in cases B2 and B4 respectively.
Figure 1 .
Figure 1.Best fit for the real data of the averaged nutrients and biofilm thickness dynamics predicted by the 1D model.Red line corresponds to the numerical results whereas the black points correspond to the experimental measurements. | 1,453.2 | 2018-09-18T00:00:00.000 | [
"Biology",
"Mathematics",
"Environmental Science"
] |
In-plane hyperbolic polariton tuners in terahertz and long-wave infrared regimes
One of the main bottlenecks in the development of terahertz (THz) and long-wave infrared (LWIR) technologies is the limited intrinsic response of traditional materials. Hyperbolic phonon polaritons (HPhPs) of van der Waals semiconductors couple strongly with THz and LWIR radiation. However, the mismatch of photon − polariton momentum makes far-field excitation of HPhPs challenging. Here, we propose an In-Plane Hyperbolic Polariton Tuner that is based on patterning van der Waals semiconductors, here α-MoO3, into ribbon arrays. We demonstrate that such tuners respond directly to far-field excitation and give rise to LWIR and THz resonances with high quality factors up to 300, which are strongly dependent on in-plane hyperbolic polariton of the patterned α-MoO3. We further show that with this tuner, intensity regulation of reflected and transmitted electromagnetic waves, as well as their wavelength and polarization selection can be achieved. Our results can help the development of THz and LWIR miniaturized devices.
The discovery of two-dimensional (2D) vdW crystals has opened avenues for exploring functional materials and devices in the THz (30-3000 μm) and long-wave infrared (LWIR; 8-15 μm) spectral ranges [1][2][3][4][5][6][7][8][9][10] . The THz and LWIR technologies are of great significance for future photonic and optoelectronic applications, such as 5 G/6 G mobile net-works 11,12 , night vision 13 , biomedical imaging and sensing 14,15 , thermal management 16 , and deep-space exploration 17 . However, their development is always limited by the scarcity of materials with strong and tunable intrinsic optical responses, in particular those used for devices in nanoscale and of room temperature operation.
In the past decades, much effort has been devoted to develop narrow band-gap semiconductors (e.g., mercury cadmium telluride and InSb) and quantum materials (e.g., quantum wells/dots, super lattice) [18][19][20] , whose inter-band, intra-band, or inter-subband optical transitions are found in the LWIR and THz regimes. However, due to the relatively weak light-matter operation, their optical responses are weak, and this is further affected by thermal noise. Thus their devices usually require a cryogenic operation to suppress thermal noise, and the introduction of components such as antennas and/or light absorbing layers to improve the electromagnetic absorption, the dimension of which will be out of the scale of tens of micrometers, not to mention the nanoscale. These will result in complex and largevolume architectures that are not favorable for nanodevices, even miniaturized and portable devices. Moreover, the broadband-and polarization-insensitive optical transitions of these materials also restrict their applications in spectrally and polarization-selective photonic and optoelectronic devices, which are in particular interesting in a modern information society. Metamaterials and metasurfaces comprised of artificially designed metallic or dielectric unit cells are able to confine the THz and LWIR electromagnetic waves to enhance light−matter interactions, which therefore give rise to a variety of functional devices [21][22][23][24][25][26] . However, despite intense research efforts, in these spectral ranges, the unit cells of many conventional metamaterial/metasurface often offer restricted confinement due to high losses 21,27 . These are not conducive to device integration, and will also increase device power consumption.
However, so far, the demonstration of utilizing the above exotic characteristics in a practical device is not given. The challenge lies in that one needs to compensate for the large photon−polariton momentum mismatch for far-field excitation and far-field characterization of the HPhPs. In the previous studies, the HPhPs of vdW crystals have been observed by near-field nano-imaging techniques [8][9][10]28 , relying on using a metallic nanotip to compensate for the large momentum mismatch between free-space photons and polaritons. For most practical device applications, direct excitation of the HPhPs from the far-field is necessary. Some earlier studies indicate that it is possible to pattern the surface of vdW crystals, such as with graphene [40][41][42] , hexagonal boron nitride 7,43,44 , semi-metals 45 , and topological insulators 46 to excite and measure the various types of polaritons. However, these previous studies focused on plasmon polaritons with in-plane isotropic [40][41][42]46 and hyperbolic dispersions 45 , as well as phonon polaritons with in-plane isotropic dispersion 7,43,44 . Far-field excitation and characterization of the tunable in-plane HPhPs in vdW crystals, especially in the THz spectral regime, remain unexplored. Furthermore, exploring the applications of the in-plane HPhPs in optical devices has so far remained elusive. These can be done if one has a vdW crystal with a large enough lateral size while maintaining thicknesses of nanometer scale, so that patterns are made larger than the diffraction limit for these free-space wavelengths and, thus, suitable for far-field spectroscopy.
In this article, we demonstrate far-field excitation and far-field characterization of HPhPs in an In-Plane Hyperbolic Polariton Tuner, which is formed by patterning one-dimensional (1D) ribbon array directly onto the semiconducting HPhP vdW α-MoO 3 flake with a centimeter lateral size while maintaining thicknesses of 100~200 nm. The THz and LWIR photons from far-field illuminating onto the tuner will strongly couple with the phonons and give rise to polaritons with in-plane hyperbolicities. This makes an in-plane tuner an actual device that acts not only with functions of grating but also as a polarizer and notch filter in the LWIR and THz regimes, and have distinctive features including high-Q (300) resonance and extinction ratios up to 6.5 dB at a deep sub-wavelength thickness of 200 nm. Moreover, the polariton resonance frequency, i.e., the operation frequency of the polarizers and notch filters, can be highly tuned by varying the period and the skew angle of the ribbon array.
Fabrication of tuners and far-field HPhPs excitation
An in-plane tuner will have a structured surface written with desirable patterns. In this study, the in-plane tuner consists of simple onedimensional periodic ribbon patterns (1D-PRPs) directly formed on a vdW α-MoO 3 flake using electron-beam lithography (EBL). They have widths (w) and skew angles (θ), which is defined by the angle between the long-axis ribbon direction and [001] crystallographic axis of vdW α-MoO 3 (Fig. 1a). The ribbon period (Λ) is set as 2w. Both of w and Λ are much smaller than the excitation wavelength. As such, we synthesized a 120 nm-thick α-MoO 3 crystal with a largest lateral size larger than 1 cm ( Fig. 1b and Supplementary Fig. 1, and see Methods for details), which guarantees the fabrication, characterization, and comparison of different tuners on the same flake ( Fig. 1c and Supplementary Fig. 2). The basal plane of the as-grown α-MoO 3 is (010) plane, with the two orthogonal directions corresponding to [100] and [001] crystallographic axes, respectively [8][9][10] . In our study, these two axes are defined as the x-and y-axes, respectively (Fig. 1a), which are identified experimentally using micro-Raman spectroscopy ( Supplementary Fig. 1b-d). A broadband THz and LWIR light illuminates the tuner and the optical responses at the far-field were measured using a polarized Fourier transform infrared (FTIR) micro-spectroscopy (Fig. 1a) (see Methods for details). It should be noted that usually, three main techniques are employed for determining the broadband polaritonic properties of 2D crystals, including the FTIR 36,47,48 , electron energy loss spectroscopy (EELS) 49 , and infrared nanoscopy 4,9,28 . In comparison with the latter two techniques, which usually require complex instrumentation, harsh sample preparation, and time consumption, the farfield polariton characteristics of the 1D-PRPs can be readily measured in a common broadband FTIR spectrometer at ambient conditions, with low time-consuming, a high collection efficiency, and over a large sample area.
In vdW α-MoO 3 crystal, in the spectral regimes 230−400 cm −1 (THz) and 545−1010 cm −1 (LWIR), there are a series of Reststrahlen bands where the Re(ε) along one of the three crystallographic axes, i.e., [100], [001], and [010], is negative ( Supplementary Fig. 3), while at least one is positive. This makes α-MoO 3 a natural hyperbolic medium capable of supporting HPhPs. We first characterized the far-field reflection of the homogeneous pristine α-MoO 3 , which is calculated as R/R 0 − 1, with R and R 0 the reflectance of the light from the surfaces of the sample and bare substrate (see Methods). For incident light polarized along the [001] and [100] directions, the reflectance spectra show distinct peaks at 550 and 820 cm −1 (Fig. 1d), which correspond to IR-active TO phonon modes along [001] ðω 001 TO Þ and [100] axes ðω 100 TO Þ, respectively. Both of the spectra exhibit small valleys at 1004 cm −1 . These valleys are very close to the frequency of LO phonon mode along the [010] axis (z-axis, ω 010 LO ) where the permittivity diminishes. For a 120 nm-thick α-MoO 3 flake, leaky modes (Berreman modes) can be excited near this epsilon-near-zero (ENZ) region and then give rise to the two valleys on the reflectance spectra 50,51 . These narrow leaky modes can further interfere with the broad reflection background and generate asymmetric Fano lineshapes (see Supplementary Note 1 and Supplementary Fig. 4). However, no evident spectral peaks corresponding to HPhPs are observed in the rest of the spectral range. This is due to the large wavevector mismatch between free-space photons (k 0 ) and polaritons (q PhPs ), which prevents the coupling of electromagnetic fields to HPhPs.
A tuner comprised of 1D-PRPs ( Fig. 1c and Supplementary Fig. 2) is able to overcome the large momentum mismatch and excite the HPhPs 52,53 . The incident waves will be scattered by the sharp ribbon edges into evanescent waves with large momenta, whereby HPhPs propagating transverse to the ribbons are excited. Fabry−Pérot resonances (FPRs) can then be formed upon the multiple polariton reflections from the ribbon edges. Simultaneously, the 1D-PRPs can also diffract the incident light into guided waves propagating perpendicular to the ribbon long axis, whose wavevectors are much larger than the free-space waves 54 . These guided waves can then couple with and transfer energy to the polariton FPRs (see Supplementary Note 2 and Supplementary Figs. 5, 6 for more discussion on the excitation of HPhPs by the 1D-PRP). The polariton energy will be dissipated by the lattice vibrations or radiated back to the free space, as manifested by resonance peaks and valleys in the corresponding reflectance and transmission spectra, respectively.
For a typical 1D-PRP with w = 800 nm and orientated along the [001] axis (sample with skew angle θ = 0°shown in Fig. 1c), HPhPs in Reststrahlen Band 2 (820 to 972 cm −1 , where Re(ε x ) <0 and Re(ε y ), Re(ε z ) >0) will be excited upon illumination polarized along the [100] axis. This will lead to a strong reflectance peak at 874 cm −1 , as shown in Fig. 1e. The small bump at 820 cm −1 originates from the intrinsic ω 100 TO . No resonance peaks are observed when the polarization is switched to [001] direction. In contrast, for 1D-PRP orientated along the [100] axis (sample with θ = 90°shown in Fig. 1c), a clear peak at 689 cm −1 is identified (Fig. 1f), suggesting the launching of HPhPs in Reststrahlen Band 1 (545 to 851 cm −1 , where Re(ε y ) <0 and Re(ε x ), Re(ε z ) >0). For polarization along the [100] axis, only the ω 100 TO peak appears. The excitation of HPhPs in both Bands 1 and 2 using these two types of 1D-PRPs can be further confirmed by simulating the near-field distributions they support. The simulated field distributions at ω = 874 and 689 cm −1 clearly reveal that polaritonic rays with zig−zag shape patterns propagate inside the ribbons ( Supplementary Fig. 7a, b and Supplementary Note 3). These are typical fingerprints of HPhP waveguide modes. These modes are bulk modes with electromagnetic fields confined inside the body of the flakes, such as those observed in hBN nanostructures 55 and biaxial α-MoO 3 flakes 56 .
Notably, in the two 1D-PRPs two valleys with narrow linewidths appear around 1000 cm −1 for both polarization conditions. These valleys are spectrally close, but occur with different amplitudes, with the spectral valleys deeper when the incident light is polarized perpendicular to the ribbon. In addition, as discussed below, the deeper valleys shift when w and θ change. Therefore, they are ascribed to HPhP resonances in Reststrahlen Band 3 (958 to 1010 cm −1 , where Re(ε z ) <0 and Re(ε x ), Re(ε y ) >0). This is also corroborated by the corresponding near-field distributions showing polaritonic rays with zig−zag shapes ( Supplementary Fig. 7c, d). Due to their narrow linewidths, Fano interference will occur between the background reflectance and the HPhP resonances, giving rise to these valley features (see Supplementary Note 1 and Supplementary Fig. 4). The shallower valleys are contributed by the aforementioned ENZ condition that occurs near the ω 010 LO (Fig. 1d). These results are consistent with polariton propagation in the basal plane of an α-MoO 3 flake: the HPhPs in Band 1 and 2 are of in-plane hyperbolicity, which cannot propagate along [100] and [001] directions, respectively 10 . In contrast, the dispersion of HPhPs in Band 3 are elliptical, which therefore allows them to propagate along both of the two orthogonal crystallographic directions 10 .
Tuning the HPhPs with tuners of different ribbon widths
The above results clearly prove that the HPhPs supported by α-MoO 3 , which previously were only accessed using near-field nanoimaging [8][9][10]30,31 , are able to be excited from the far-field using the 1D-PRPs. To demonstrate the tunability of HPhPs in this material, we measured the reflectance spectra of 1D-PRPs with different w ranging from 100 to 2000 nm. All spectra were collected with the incident light polarized perpendicular to the ribbon long axis. For 1D-PRPs parallel to the [100] direction, both of the HPhP resonances corresponding to Band 1 (peaks) and 3 (valleys) can be excited (Fig. 2a). Notably, when w increases from 600 to 2000 nm, the resonance in Band 1 clearly redshifts from 746 to 613 cm −1 , whereas a reverse trend appears for the modes in Band 3, where the resonances blueshift from 998 to 1005 cm −1 . Similar spectral evolution with changing w can be observed for 1D-PRPs along the [001] direction (Fig. 2b), where PhP resonances in Band 2 and 3 are excited. Additionally, the variation of resonance frequency in Band 3 with w is different for these two ribbon orientations ( Supplementary Fig. 8).
The evolution of the HPhP resonances with changing w can be understood by considering that the excitation of HPhPs are originated from the synergy between guided waves of the array and FPRs in an individual ribbon: the scattering of light at the ribbon edges excite the polariton FPR, while the guided waves further couple with and transfer energy to the polariton waves (see Supplementary Note 2 and Supplementary Fig. 5, 6). The conditions for the occurrence of the FPRs and guided waves are q PhPs w = ± mπ and q PhPs Λ = ±n2π, respectively, with m, n = 1, 2, 3, 45,57 . Because in our study, the Λ is deliberately set as 2w, these two equations are the same. Therefore, each w corresponds to an in-plane polariton wavevector of q PhPs = ∓ mπ w . When w changes the resonance peak will scale according to the in-plane polariton dispersion relations, ω(q PhPs ). This can be readily seen by calculating the ω(q PhPs ) along the [100] and [001] directions, which is visualized as the 2D false color plot of the imaginary part of the complex reflectivity Imr p (q PhPs , ω) (see Supplementary Note 4 for details on the calculation of Imr p ). The PhP resonance frequencies obtained from the spectra shown in Fig. 2a, b are then overlaid onto the same plot by using q PhPs = mπ/w. Excellent agreement is obtained between the experimental measurements and calculated lowest-order (m = 1) HPhP branches for all three bands (Fig. 2c). Such an agreement further validates that the highly anisotropic HPhPs in the α-MoO 3 flake are directly excited from the far-field with the help of the in-plane hyperbolic polariton tuners. It is noted that the reflection of polariton waves by the ribbon edges may induce phase shifts, which can violate the condition for the FPRs by a phase of 2Φ 58 . A previous theoretical result showed that in monolayer graphene, the Φ for plasmon polaritons is0 .75π 58 . In comparison with the plasmon polariton in graphene, the propagation and reflection of the in-plane HPhPs are rather complicated, making the phase shift difficult to be predicted. In our analyses, the Φ is taken as π according to a very recent study 59 . The good agreement between the experimental measurements and calculated results further validates our setting.
The far-field reflectance spectra allow for evaluating the Q-factor of the HPhP resonance, which is defined as Q = ω 0 Γ , with ω 0 and Γ, the frequency and linewidth of a specific resonance 7 (see details in Supplementary Note 1 for extraction of the Q-factors). For HPhP resonances in Band 1 and 3, their Q-factors respectively decrease and increase monotonically against ω 0 (Fig. 2d). This is because when ω 0 increases, the PhP resonance in Band 1 shifts closer to the ω 100 TO (820 cm −1 ), while the resonance in Band 3 shifts away from the TO phonon mode along [010] axis at 958 cm −1 ðω 010 TO Þ 10,36 . Thus the polariton dissipation by lattice absorption will be strengthened (suppressed) for Band 1 (Band 3), giving rise to a larger (smaller) Γ. The non-monotonic behavior of the Q-factor for the HPhP resonance observed in Band 2 can be understood by considering that, in addition to ω 100 TO , there is another LO phonon mode along the [100] axis at 972 cm −1 ðω 100 LO Þ 10,36 . Leaky-mode absorption induced by ENZ condition also occurs near the ω 100 LO . Therefore, the Q-factor first increases as ω 0 is farther from the ω 100 TO , and then decreases gradually approaching the ω 100 LO . It is noted that the Q-factors observed in Band 1, 2, and 3 are 15-25, 25-100, and 200-300, respectively. Most of these values are higher than those observed in graphene nano-gratings with similar resonance frequencies 41,42,47 . In particular, the Q-factor of the Band 3 resonance can be as high as 300, which is on par with the highest observed in hBN nanoresonators (360) via the same far-field technique 60 . Such high Qfactors, coupled with the small modal volumes and footprint of the 1D-PRPs, indicate that the α-MoO 3 tuners offer important application potential for high-efficiency compact photonic devices and components, as demonstrated below.
The far-field spectra also allow for extracting the polariton lifetimes, which span from 0.2 to 3.0 ps in the three Reststrahlen bands (Supplementary Note 5, Supplementary Fig. 9, and Supplementary Table 1). It is noted that the fabrication of 1D-PRPs can lead to damage to the ribbon edges, which can reduce the polariton lifetimes and Q factors. Usually, this issue is inevitable during the patterning of the vdW crystal for various characterizations and device applications. To evaluate the additional polariton damping induced by the patterning processes, we performed near-field measurements on the same α-MoO 3 flake in the unpatterned region and extracted the intrinsic polariton lifetimes (Supplementary Note 5, Supplementary Fig. 9, and Supplementary Table 1). In comparison with the unpatterned α-MoO 3 , the phonon polariton lifetime in the ribbon arrays are reduced by 20−57%. The reduction in the lifetime of polariton modes in α-MoO 3 ribbons is attributed to defects or impurities at the rough edges of the ribbons introduced during the fabrication process. These imperfections can create additional scattering centers for HPhPs. Furthermore, they will also lead to the broadening of the pristine phonon modes. Raman spectroscopic characterizations show that, compared to the unpatterned region, the overall linewidths of the phonon modes in the ribbons increased by 4.6-14.1% (Supplementary Note 6, Supplementary Fig. 10, and Supplementary Table 2). This broadening shortens the lifetime of the associated polariton modes as well. Despite the lifetime reduction, the 1D-PRPs still exhibit high Q factors upto 300. Although such a value is smaller than that of resonances sustained by an unpatterned and naturally grown α-MoO 3 ribbon 61 , it can be further improved by optimizing the processing parameters.
Tuning the HPhPs with ribbon arrays of different skew angles
The in-plane HPhP dispersions of the α-MoO 3 are highly anisotropic 9,10 . This offers unique tunability of the HPhP resonances by changing the 1D-PRP orientations, which cannot be realized in vdw crystals with isotropic in-plane dispersions such as hBN and graphene. As such, 1D-PRPs with fixed w (480 nm), but different θ were fabricated (Figs. 3a, 1c). Each pattern can provide polariton momenta of q PhPs = π/w, with the direction perpendicular to the ribbon long axis. In this way, when the ribbon is rotated away from the [001] axis, HPhPs with wavevectors of different orientations within the basal plane can be excited, giving rise to HPhP resonances that are strongly dependent on the θ. Specifically, HPhP resonances in Band 2 and 3 can be observed for θ = 0°, and all of the HPhP resonances in Band 1, 2, and 3 appear when θ is increased (Fig. 3b). For θ = 90°, the resonance in Band 2 merges with the ω 100 TO . In addition, the resonances in Band 1 and 2 are highly sensitive to θ, while that in Band 3 shifts slowly against θ (Fig. 3b). This can be ascribed to the distinct in-plane polariton dispersions of the three bands. Specifically, the in-plane isofrequency contour (IFC) of HPhPs in Band 3 is an ellipse, where the polariton dispersion differs moderately along different directions in the basal plane. However, the IFCs in Band 1 and 2 are hyperbola, making their polariton dispersions highly dispersive with θ. These angle-dependent behaviors can be seen more clearly by plotting the dependence of resonance frequencies in the three bands on the skew angle (Fig. 3c), which agrees well with the calculated Imr p (q PhPs , ω). The HPhP dispersion relations at each skew angle can be obtained by measuring the reflectance spectra from 1D-PRPs with different w (q PhPs ) at a specific θ ( Supplementary Fig. 12), whereby the in-plane polariton IFCs at different energies can be re-constructed and visualized. Clearly, the HPhP resonances in Band 1 and 2 depict IFCs of open hyperbolic shapes (Fig. 3d, e), while those in Band 3 correspond to an IFC of a closed ellipse (Fig. 3f). Moreover, at higher frequencies in Band 1 (Band 2), the opening-angles of the hyperbolic sectors become smaller and the hyperbola bends toward the [001] ([100]) direction (Fig. 3d, e). All the experimental points can be fit well by the calculated Imr p (pseudo-colored plots shown in Fig. 3d-f and Supplementary Fig. 12). These results provide further direct evidence for far-field excitation and modulation of the hyperbolic HPhPs in the α-MoO 3 flake.
High-Q resonances can also be induced in 1D-PRPs made out of an in-plane isotropic polaritonic vdW crystal 44 . The frequency of such resonances can also be tuned by changing the ribbon width, but the α-MoO 3 1D-PRPs proposed in the current study is unique. The in-plane hyperbolicity of α-MoO 3 makes it possible to tune the resonance frequency of the 1D-PRPs by rotating the ribbons while fixing their widths. This feature can greatly simplify the fabrication processes of arrays with different resonance frequencies, and small additional damping will be introduced because the ribbon width is unchanged. Moreover, for a polariton mode with a wavevector approaching the asymptote of the hyperbolic IFC, the polariton momentum will become remarkably high. This will generate much stronger electromagnetic field confinement than those with wavevectors away from the asymptote. These modes can significantly enhance light−matter interactions at the nanoscale and lead to various applications, such as enhanced light emission, ultrasensitive biosensing, and nonlinear optical signal generations. For these applications, usually, a fixed operation frequency is preferred. In the 1D-PRPs made out of α-MoO 3 , tuning the orientation and width of the ribbons can both tailor the resonance frequency. Therefore, it is possible to induce a high-momentum polariton mode while fixing its resonance frequency by simultaneously orientating the ribbon long axis along the asymptote of the hyperbolic IFC and tuning the ribbon width. This feature can open an avenue for the applications just mentioned.
Far-field excitation of tunable THz HPhPs with ribbon arrays
The α-MoO 3 flake can also support nanoscale-confined HPhPs in the THz domain from 260-400 cm -1 (8-12 THz), which, however, have only been probed using the near-field nano-imaging technique 30 . To demonstrate the far-field excitation and tuning of the THz HPhPs, we fabricated 1D-PRPs of different w and θ and characterized their spectral responses. Due to the relatively low signal-to-noise ratio of the bolometer in the THz domain, transmission spectra were recorded, which is defined as T/T 0 , with T and T 0 the transmittance of the light through the sample and bare substrate (Methods (Fig. 4a). The tunability of these two resonances is clearly demonstrated by their redshifting behaviors with increasing w. The polariton dispersion relations were then obtained by extracting the resonance frequencies at different w. The analytical dispersions were derived using the dielectric tensor reported in ref. 30 and are in good agreement with the experiment measurements ( Fig. 4b and Supplementary Note 7). The Q-factors of the two types of resonances can be evaluated according to the transmittance spectra shown in Fig. 4a, which both increase with increasing the resonance frequency (Fig. 4d). The available Q-factors are in the range of 15-25 (HPhP [001] ) and ) respectively, which surpass that of graphene plasmon resonance in the THz regime 40 . The HPhP resonances are also strongly dependent on the skew angle of the 1D tuner patterns. By sweeping the θ from 0°to 90°( ribbon long axis rotating from [001] to [100] direction), the HPhP [001] and HPhP [100] resonances shift monotonically to higher and lower frequencies, respectively (Fig. 4c). Moreover, with the experimental data shown in Fig. 4b, c, the in-plane IFC of the HPhP [001] can readily be drawn to exhibit a clear hyperbola opening towards the [001] axis ( Supplementary Fig. 13). These results corroborate the excitation of HPhP resonances in THz Reststrahlen Band 1 (HPhP [001] ) and 3 (HPhP [100] ), which are consistent with the previous nano-imaging results 30 .
Tunable LWIR and THz polarization notch filters
A PNF is a unique optical component that combines a polarizer and a narrow band-rejection filter together into a single component, which is able to block a monochromic laser with a given linear polarization, while passing light of all polarization states at wavelengths adjacent to the laser line. Such a filter has broad application prospects in laser spectroscopy and optical communications, but the commercial products are rare, and especially, there is no commercial PNF in the LWIR and THz ranges. The salient high-Q (Figs. 2d, 4d) and polarizationsensitive (Fig. 1e, f) α-MoO 3 tuners established above provide opportunities for developing tunable PNFs 62 . As such, we constructed PNFs using 1D-PRPs with long axes along [001] and [100] axes, respectively. For typical 1D-PRPs with w = 1000 nm, their extinction spectra in LWIR (Fig. 5a) and THz (Fig. 5b) regimes are strongly polarizationdependent. Specifically, the HPhP resonances only appear when the excitation polarization is perpendicular to the ribbon's long axis. This can be seen more clearly by plotting the extinction at the corresponding resonance peaks, i.e., 650 cm −1 /270 cm −1 and 869 cm −1 / 362 cm −1 , against light polarization for ribbons parallel to [100] and [001] directions, respectively (Fig. 5c, d). The performance of a PNF for blocking a monochromatic light source can be evaluated by two parameters: the polarization extinction ratio and bandwidth. Specifically, the polarization extinction ratio is defined as 10 log(T 0 /T), with T 0 and T the transmittance of light polarized along and perpendicular to the ribbon's long axis at the resonance frequency. The bandwidth can be calculated according to the full width at half maximum (FWHM) of a specific resonance. Accordingly, for the PNFs with resonances at 650, 869, 270, and 362 cm −1 , corresponding extinction ratios (bandwidth) are 3.4 dB (41.5 cm −1 ), 5.5 dB (26.5 cm −1 ), 4.5 dB (23.4 cm −1 ), and 3.7 dB (24.5 cm −1 ), respectively. In particular, the bandwidths of the PNFs are comparable to many of the commercial narrow-band-pass filters in similar spectral regimes, which are usually polarization insensitive, (Supplementary Fig. 14), while the thicknesses of the PNFs (200 nm) are much smaller than those of the commercial components (~1 mm). More importantly, using 1D-PRPs with different w, the operational frequency, polarization extinction ratio, and FWHM of the PNF can be engineered continuously ( Supplementary Figs.15, 16). The maximum peak extinction can reach 6.5 dB and the smallest FWHM can be as narrow as 17 cm −1 .
Discussion
We have successfully demonstrated direct far-field excitation and characterization of the tunable LWIR and THz HPhPs in biaxial vdW α-MoO 3 patterned into simple 1D-PRPs. The 1D-PRPs can act as polariton tuners that are sensitive to the excitation polarization and with light extinction ratios up to 6.5 dB and high Q-factors up to 300. Such a compositional set of output functions are tunable and strongly dependent upon the in-plane hyperbolic phonon polaritons in the α-MoO 3 . It is noted that in comparison with the recently reported patterned vdW WTe 2 flakes with tunable in-plane hyperbolic plasmon polaritons at cryogenic conditions 45 , the polariton tuners proposed in the current study may be more favorable for practical applications because of their room-temperature operation, broader spectral range, and much higher quality factors of the resonance modes.
On the prospects, these in-plane hyperbolic polariton tuners can be used in optical circuitry, instruments and even modern information systems. The in-plane hyperbolic polariton tuner also opens up avenues for a variety of practical photonic and optoelectronic applications besides the PNFs shown in Fig. 5. For example, by engineering the tuner to spectrally overlap the HPhP resonance with a specific vibration or rotation transition of a molecule, strong interactions between the tuner and molecule can be induced 44,63 , which can significantly enhance the molecular absorption or emission and give rise to various ultrasensitive bio-sensing techniques. The tuners with high-Q and tunable HPhP resonances can also be employed to regulate the blackbody emission 62 , whereby narrow-band, polarized, and tunable thermal emission can be achieved [64][65][66] . Moreover, tunable and highperformance LWIR and THz photodetectors can also be envisioned by taking advantage of the strong light field localizations ( Supplementary Fig. 7) and semiconductor nature of the α-MoO 3 19 ; this type of devices may give unique functions necessary for future communication and radar applications.
For fundamental research, the far-field excitation methodology can complement near-field nano-imaging techniques and make the characterizations of PhPs of materials more precisely, especially for those with in-plane hyperbolicity. For example, in nano-imaging measurement, because the polariton waves are launched by an antenna (the scanning tip or antenna fabricated onto the sample surface), in principle, these HPhPs can propagate along different directions determined by the hyperbolic IFC [8][9][10] . Therefore, the measurements of the polariton wavelength and dispersion relation from the polariton interference fringes can be disturbed by these various polariton waves. On the other hand, with far-field excitation, the polariton wavevector is determined by the tuner's structural parameter and orientation. For a specific pattern, only one HPhP mode can be excited, which allows for a more accurate characterization of the intrinsic HPhP properties. Additionally, it is possible to characterize the HPhPs using a broadband light source covering a broad THz spectral range. This can unveil more complete polaritonic properties for broadening and deepening our understanding of the THz HPhPs, in particular for two-dimensional atomic crystals (e.g., the in-plane IFCs in the THz domain as shown in Supplementary Fig. 13), which is now limited by the discrete laser lines used in most nano-imaging measurements 30 . It is noted that in the current study, as a demonstration of principle, we only employ the simplest form of patterns, and only demonstrate with one type of material. In fact, patterns can be chosen depending on desirable applications, and also materials as long as supporting HPhPs in the THz and LWIR regimes 6,67 . The tuner allows us to modulate the wavefront of the incident light and control their power flow in an engineered space [21][22][23][24][25][26] , which therefore enables a variety of interesting applications, such as negative refraction, holography, metalens, polarization conversion, and even topological PhPs and exciton-polaritons with robust beam steering functionalities 68,69 , in visible and near-infrared ranges to be expanded into the THz and LWIR spectral regimes.
When preparing the revised manuscript, we became aware of a recent work 59 on a similar topic to our current study.
Methods
Synthesis of large-area vdW α-MoO 3 flake vdW α-MoO 3 thin flakes with a centimeter lateral size were prepared using a modified thermal physical vapor deposition method. Specifically, an alumina crucible filled with 0.1-g α-MoO 3 powders was placed at the center of a quartz tube as the evaporation source. Another crucible covered with a silicon substrate of 1 cm × 1 cm was placed 20 cm away from the source. The source was then heated up to 780°C [001] and held at that temperature for 2 h. The α-MoO 3 powders were sublimated and crystallized onto the silicon wafer. Afterward, the quartz tube was cooled down to room temperature naturally. The large-area α-MoO 3 flakes can be found on the silicon wafer.
Fabrication of the vdW α-MoO 3 1D periodic tuner patterns
The as-grown α-MoO 3 flakes were transferred to a pristine (highly resistive) silicon substrate covered with a 300-nm oxide layer. Afterward, a selected flake was patterned into 1D-PRPs with different w and θ using a combination of electron beam lithography (EBL: EBPG5000+, Netherlands) and reactive ion etching (RIE: 50 W for 10 min). For the EBL processing, a 400-nm layer of Poly(methylmethacrylate) (PMMA) photoresist was used. For the RIE etching, a mixture of O 2 (12 vol.%), Ar (30 vol.%), and CHF 3 (58 vol.%) was employed. The etching was conducted at 50 W for 10 min. To guarantee good signal-to-noise ratios of the spectral characterizations, the areas of the patterns were set as 50 μm × 50 μm and 300 μm × 300 μm for LWIR and THz regimes, respectively.
Reflectance and transmittance spectral characterizations LWIR and THz spectral characterizations were performed using a Bruker FTIR spectrometer (Vertex 70 v) integrated with a Hyperion 3000 microscope and a mercury cadmium telluride (HgCdTe) photoconductor (for the measurement of LWIR spectra) or a liquid-helium-cooled silicon bolometer (or the measurements of THz spectra) as the detector. A broadband black-body light source covering the LWIR and THz spectral regimes was employed as the incident light. A 15× reflective Schwarzschild objective was utilized to focus the incident light onto the tuner patterns with a spot size of~500 μm. The reflected or transmitted light were collected from an area of 50 μm × 50 μm and 300 μm × 300 μm for LWIR and THz regimes, respectively, with the help of an iris diaphragm. For the polarization-dependent measurements, a polarizer was used to control the polarization of the incident light. For the reflectance and transmittance spectra of the tuner, a bare silicon substrate is used as reference for normalization.
To characterize the PNFs in LWIR and THz regimes, a linear polarizer was placed before the detector to determine the polarization state of light transmitting through the tuners upon an unpolarized illumination.
Data availability
Relevant data supporting the key findings of this study are available within the article and the Supplementary Information file. All raw data generated during the current study are provided in the Source Data file or available from the corresponding authors upon request. Source data are provided with this paper.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. | 8,597.4 | 2022-06-21T00:00:00.000 | [
"Physics"
] |
The Prognostic Value of ASPHD1 and ZBTB12 in Colorectal Cancer: A Machine Learning-Based Integrated Bioinformatics Approach
Simple Summary Colorectal cancer (CRC) is among the leading causes of cancer-related deaths. Despite extensive efforts, a limited number of biomarkers and therapeutic targets have been identified. Therefore, novel prognostic and therapeutic targets are needed in the management of patients and to increase the efficacy of current therapy. The majority CRC patients follow the conventional chromosomal instability (CIN), which is started by several mutations such as APC, followed by genetic alterations in KRAS, PIK3CA and SMAD4, as well as the hyperactivation of pathways such as Wnt/TGFβ/PI3K. Although the underlying genetic changes have been well identified, the mutational signature of tumor cells alone does not enable us to subclassify tumor types or to accurately predict patient survival and suppression of those pathways have often not been effective in treatment. Our data showed some new genetic variants in ASPHD1 and ZBTB12 genes, which were associated with a poor prognosis of patients. Abstract Introduction: Colorectal cancer (CRC) is a common cancer associated with poor outcomes, underscoring a need for the identification of novel prognostic and therapeutic targets to improve outcomes. This study aimed to identify genetic variants and differentially expressed genes (DEGs) using genome-wide DNA and RNA sequencing followed by validation in a large cohort of patients with CRC. Methods: Whole genome and gene expression profiling were used to identify DEGs and genetic alterations in 146 patients with CRC. Gene Ontology, Reactom, GSEA, and Human Disease Ontology were employed to study the biological process and pathways involved in CRC. Survival analysis on dysregulated genes in patients with CRC was conducted using Cox regression and Kaplan–Meier analysis. The STRING database was used to construct a protein–protein interaction (PPI) network. Moreover, candidate genes were subjected to ML-based analysis and the Receiver operating characteristic (ROC) curve. Subsequently, the expression of the identified genes was evaluated by Real-time PCR (RT-PCR) in another cohort of 64 patients with CRC. Gene variants affecting the regulation of candidate gene expressions were further validated followed by Whole Exome Sequencing (WES) in 15 patients with CRC. Results: A total of 3576 DEGs in the early stages of CRC and 2985 DEGs in the advanced stages of CRC were identified. ASPHD1 and ZBTB12 genes were identified as potential prognostic markers. Moreover, the combination of ASPHD and ZBTB12 genes was sensitive, and the two were considered specific markers, with an area under the curve (AUC) of 0.934, 1.00, and 0.986, respectively. The expression levels of these two genes were higher in patients with CRC. Moreover, our data identified two novel genetic variants—the rs925939730 variant in ASPHD1 and the rs1428982750 variant in ZBTB1—as being potentially involved in the regulation of gene expression. Conclusions: Our findings provide a proof of concept for the prognostic values of two novel genes—ASPHD1 and ZBTB12—and their associated variants (rs925939730 and rs1428982750) in CRC, supporting further functional analyses to evaluate the value of emerging biomarkers in colorectal cancer.
Introduction
Colorectal cancer (CRC) is the second most common cause of cancer-related mortality [1], and its incidence is increasing despite the advances in the detection of prognostic and/or therapeutic targets.This is partly due to the limited number of therapeutic agents that have been identified.A high proportion of patients with CRC develop metastatic cancer(s) or become resistant to therapy.Therefore, novel prognostic biomarkers and new therapeutic targets that can help to assess the risk of developing CRC recurrence or increase the efficacy of current therapy are urgently needed.
Integrated analyses of multi-omics data provide useful insight into the pathogenesis of CRC and help to identify novel diagnostic and prognostic biomarkers.With the success of artificial intelligence technologies, machine learning (ML) is being used in healthcare.ML methods provide novel techniques of integration and analyzing omics for the discovery of novel biomarkers [2,3].Hammad and collaborators [4] identified 105 differential expression genes (DEGs) using datasets from the Gene Expression Omnibus (GEO).Functional enrich-ment analysis revealed that these genes were enriched in cancer-related biological processes.The protein-protein interaction (PPI) network selected 10 genes, including IGF1, MYH11, CLU, FOS, MYL9, CXCL12, LMOD1, CNN1, C3, and HIST1H2BO, as hub genes.Support Vector Machine (SVM), Receiving Operating Characteristic (ROC), and survival analyses demonstrated that these hub genes can be considered potential prognostic biomarkers for CRC.
Maurya et al. [5] used Least Absolute Shrinkage and Selection Operator (LASSO) and Relief for feature selection from the Cancer Genome Atlas (TCGA) dataset and applied RF, K-Nearest Neighbor (KNN), and Artificial Neural Network (ANN) to check the accuracy of the models.The joint set of selected features between LASSO and DEGs was 38 genes, among which VSTM2A, NR5A2, TMEM236, GDLN, and ETFDH were correlated with the overall survival (OS) of patients with CRC and could be used as prognostic biomarkers.For example, Liu et al. [6] identified 16 lncRNAs as an immune-related lncRNA signature (IRLS) for predicting patients' prognosis of CRC using machine learning-based integrated analysis.They performed further investigations to validate the application of IRLS in practice.The efficacy of immune-related lncRNA signature was validated using qRT-PCR on CRC tissues collected from 232 patients.A prospective cohort study, RECOMMEND (NCT05587452), aimed to assess the accuracy of a novel AI-based integrated analysis screening method for CRC and advanced colorectal adenomas using plasma multi-omics data.
Genome-wide association studies (GWAS) have already allowed significant progress in the understanding of the complex genetics behind the pathogenesis of CRC.There are at least three major molecular pathways that can lead to CRC, including the chromosomal instability pathway (characterized by aneuploidy or structural chromosomal abnormalities), chromosomal instability, and mutations (e.g., APC, KRAS, PIK3CA, SMAD4, or TP53).There is a growing body of evidence on targeting deregulated intracellular pathways, such as the hyperactivation of WNT-β-catenin, PI3K/Akt, or RAS signaling, although it has been shown that inhibiting these pathways has often not been effective in the clinical management of CRC [7][8][9][10].Many patients with CRC had conventional chromosomal instability (CIN), which is started by several mutations such as APC, followed by genetic alterations in KRAS, PIK3CA, and SMAD4, as well as the hyperactivation of pathways such as Wnt/TGFβ/PI3K.Although the underlying genetic changes have been sufficiently identified, the mutational signature of tumor cells alone does not enable us to subclassify tumor types or to accurately predict patients' survival, and the suppression of those pathways has often not been effective in treatment [11].In this study, we attempted to develop and validate novel prognostic biomarkers based on ML-based integrated analysis as well as validation of novel candidate genes in two additional cohorts of CRC in DNA and RNA levels using whole exome sequencing (WES) and reverse transcription polymerase chain reaction (RT-PCR), respectively.
Data Sources and Data Processing
RNA-sequencing (RNA-seq) expression data and clinicopathological information were retrieved from The Cancer Genome Atlas (TCGA) database, which included 287 CRC tissue samples and 41 non-cancers tissue samples.In this study, RNA-seq data were obtained from TCGA-colorectal adenocarcinoma.Patients with colorectal cancer were classified into early-stage and late-stage.Early-stage CRCs were classified into three subgroups based on microsatellite instability (MSI) status: low MSI (MSI-L), high MSI (MSI-H), and MSI-stable (MSI-S).Late-stage CRCs were classified into two subgroups based on the therapeutic regimens (chemotherapy versus targeted therapy).
Patient's Samples
Sixty-four CRCs were included in this study based on histological confirmation by two pathologists.All the eligible patients were chemotherapeutic naive patients treated at the Omid Hospital of Mashhad University of Medical Sciences.The study was approved by the local Hospital Ethic Committee of Mashhad University of Medical Sciences.
DNA-Seq and Whole Exome Sequencing
Data from the TCGA database were downloaded and prepared for further analysis in the R programming language.The data were downloaded in the Mutation Annotation Format (MAF).MAF is a standardized format used by TCGA for storing and analyzing various types of somatic mutations in cancer.The patients were divided into two groups: patients in the early stages (I, II) of CRC and patients in the advanced metastatic stage (IV).The first group consisted of 118 patients, while the second group consisted of 28 patients.MAF data belonging to each group is analyzed with the maftools package in R programming.
The genes with a significant p-value of less than 0.05 obtained from the survival analysis were combined with the whole exome sequencing data of TCGA for colorectal cancer.Then, the variants of the candidate genes obtained from sequencing data were analyzed using the Maftools package.Then, two candidate genes, ASPHD1 and ZBTB12, were further evaluated for their impact on gene expression using RegulomeDB and 3DSNP.Subsequently, the candidate genes were further confirmed in an additional cohort performed for the Whole Exome Sequencing (WES) data of 15 patients with CRC, as described previously.
Differential Gene Expression Analysis
Normalization was performed, while the PCA plots, volcano plots, heatmap, and karyoplote were represented by the R packages "ggplot2", "heatmap", and karyoploteR to visualize data.Significance analysis of differentially expressed genes (DEGs) was performed using DESeq2 in R software with the cutoff criteria of |log fold change | ≥ 1.5 and an adjusted p-value of <0.05.
Gene Set, Ontology, and Pathway Enrichment Analysis
The significant enrichment analysis of DEGs was assessed based on Gene Ontology (GO), Reactom, GSEA, and Human Disease Ontology (DO).GO analysis (http://www.geneontology.org/) is used for annotating genes and gene products and investigating the biological aspects of high-throughput genome or transcriptome data, including biological processes, cellular components, and molecular function.The Reactom database was used for the analysis of gene functions in biological signaling pathways.We set a p-value < 0.05 and a false discovery rate (FDR) < 0.05 as the statistically significant criteria to output.The whole transcriptome was employed for GSEA, and only gene sets with p-value < 0.05 and FDR q < 0.05 were set as statistically significant criteria.Statistical significance was set at an adjusted p-value of <0.05.Several R packages were utilized to perform enrichment analyses, including ReactomePA, enrichplot, clusterProfiler, and topGO.
Survival Analysis
The univariate/Cox proportional hazards regression model was used to identify DEGs that were significantly correlated with overall survival and assess the independent prognostic factors.R version 4.2.1 software was used to analyze the data.
Machine Learning Method
Two machine learning techniques were used, including the decision tree learning and deep learning.Deep learning models were applied to identify the effective factors.The significant variables obtained from the feature selection method (Weight by Correlation) were the final parameters in creating the model.The coefficient of correlation between variables is presented as a correlation matrix.The correlation coefficient is measured from -1 to 1; positive values represent that the variables are in the same direction, and negative correlations show the variables in opposite directions.The lack of correlation was displayed as 0.
Computational Workflow
Python3.7 was utilized for modeling.Parameters of epochs = 10, activation function = Rectifier, and learning rate = 0.01 were set in deep learning.The standard workflow was utilized as follows: Splitting the source data set into a training set and test set was performed to provide some independent evaluation levels.Subsequently, the model was optimized using the training data and then independently evaluated using the test data.In this study, a 70/30 train/test ratio was determined for the ML models.For each workflow, a model with the fixed optimal hyperparameter values was retrained on data and randomly sampled from the complete dataset.Machine learning method assessment was performed by 5 indicators, including accuracy, R2, MSE, and AUC.
Accuracy = (TP + TN)/(TP + TN + FP + FN)
where TP is true positive, FP is false positive, TN is true negative, and FN is false negative.
MSE (Mean Squared Error
where Σ represents a symbol that means "sum", n is the sample size, actual is the actual data value, and the forecast is the predicted data value.R2 (R-Squared) = 1 − Unexplained Variation/Total Variation R2 is the coefficient of determination, and it tells you the percentage variation in y explained by x-variables.AUC (Area Under the Curve) represents the degree of separability and illustrates the capability of the model in distinguishing the classes.
Protein-Protein Interaction (PPI) Network
The STRING database (https://string-db.org/)was checked to find the relationship between the studied proteins obtained from DEG and the proteins that are directly or indirectly involved in the development of cancers.A minimum effective binding score of ≥0.4 was established.Genes with significant interactions were screened.
Kaplan-Meier Survival Curve
Kaplan-Meier survival curve comparison was conducted to measure the prognostic value of candidate genes in CRC using the log-rank test.
Receiver Operating Characteristic (ROC) Curve Analysis
Receiver operating characteristic (ROC) curves are a fundamental analytical tool for assessing diagnostic tests and identifying diagnostic biomarkers.ROC curve analysis evaluates the accuracy of a test to differentiate between diseased and healthy cases, thereby measuring the overall diagnostic performance [12].A ROC curve and the area under the curve (AUC) were employed to determine the specificity, sensitivity, likelihood ratios, positive predictive values, and negative predictive values using the R package (pROC, version 1.16.2).
Quantitative Real-Time-PCR Validation
Total RNAs were extracted from tissues using a total RNA extraction kit according to the manufacturer's protocol (Parstous, Tehran, Iran).RNA quantity and quality were assessed using a Nanodrop 2000 spectrophotometer (BioTek, USA EPOCH), and forty RNAs that passed the quality control were used for the next step.The RNAs were then reversetranscribed into complementary DNA (cDNA) using a cDNA Synthesis Kit (Parstous, Tehran, Iran) according to the manufacturer's instructions.Primers were designed (Forward Reverse: ASPHD1: AGTGGCTCACAATGGCTCC and AAGACAAAGTCGAGGGCCTG and ZBTB12: TTGCTCCTCTCCTGCTACACG and AACTGGCTGAGGGCATTCCG), and RT-PCR was performed via the ABI-PRISM StepOne instrument (Applied Biosystems, Foster City, CA, USA) using the SYBR green master mix (Parstous Co. Tehran, Iran).Gene expression data were standardized to glyceraldehyde 3-phosphate dehydrogenase (GAPDH) using a standard curve of cDNAs obtained from quantitative polymerase chain reaction (qPCR) Human Reference RNA (Stratagene, La Jolla, CA, USA).
Whole Exome Sequencing
The Mutation Annotation Format (MAF) data were divided into two groups: patients in the early stages and advanced metastatic stage, as shown in Figures 1 and 2, containing 118 and 28 patients, respectively.The MAF data were analyzed using the maftools package in R programming.Figures 1 and 2 show different plots, including plot maf Summary, oncoplots, Transition and Transversions reports, Plotting VAF (Variant Allele Frequencies), Somatic Interactions reports, Drug-Gene Interactions, and Oncogenic Signaling Pathways to visualize the MAF distribution in a different group.As shown in Figures 1A and 2A, in the early and late stages, missense mutations were more frequent than other mutations, and they were typically referred to as single-nucleotide polymorphism (SNP) types.Additionally, in both groups, 70-71% of patients had mutations in their APC or TP53 genes.Most of the variants are involved in Wnt/B-catenin _signaling, Genome integrity, and MAPK signaling (Figures 1B and 2B).The clonal status of the most mutated genes can be estimated using the Variant Allele Frequencies plot; clonal genes usually have an average allele frequency of about 50% in pure samples.In the early stages of tumor development, TP53 was observed to have clonal status in the tumor tissue, while SMAD4, RYR4, and TP53 exhibit such a status in the late stages, as shown in Figures 1D and 2D.Somatic Interactions analysis indicated exclusive or co-occurrence (Figures 1E and 2E).Mutually exclusive events happen in cancer when mutations in one gene prevent the occurrence of mutations in another gene.Co-occurring events, on the other hand, arise when mutations in two or more genes occur together more frequently than would be expected by chance.Determining mutually exclusive genes implies that these genes may participate in the same pathway or process, and there might be functional overlap between them.On the other hand, identifying genes that co-occur may indicate that they collaborate to facilitate the growth of tumors, or that their cumulative impact is essential for the development of cancer.The interaction between genes and drugs that target tyrosine kinase, transcription factor complex, DNA repair, and other related processes is illustrated in Figures 1F and 2F.The involvement of mutated genes in colorectal cancer across different oncogenic signaling pathways, including RTK-RAS, Wnt, Hippo, Notch, and others, is demonstrated in Figures 1G and 2G.
Gene Expression Profiling, Identification of DEGs, and Pathway Enrichment Analysis
We performed gene expression profiling in 287 CRC cases, analyzed by the DESeq2 package, according to the adjusted p-value of <0.05 and a |logFC| ≥ 1.5 (Table S1).The PCA plots, volcano plots, and heat maps of each subgroup are shown in Figures 3 and S1.Moreover, the gene expression of each subgroup, obtained from the DEG analysis was exhibited in the ideogram of chromosomes using the karyoploteR package (Figure 3C).Enrichment analysis showed that DEGs were significantly enriched in biological processes related to cancer progression.Based on GO analysis, the main biological processes involving the DEGs included ion homeostasis, inorganic cation transmembrane transport, and the regulation of hormone levels.In terms of cellular components, the DEGs were mostly enriched in the external encapsulating structure and extracellular matrix (ECM).In terms of molecular functions, the DEGs were linked by cation transmembrane transport activity, receptor regulator activity, signaling receptor activator activity, etc. (Figures 4A and S2-S6).
Gene Expression Profiling, Identification of DEGs, and Pathway Enrichment Analysis
We performed gene expression profiling in 287 CRC cases, analyzed by the DESeq2 package, according to the adjusted p-value of <0.05 and a |logFC| ≥ 1.5 (Table S1).The PCA plots, volcano plots, and heat maps of each subgroup are shown in Figures 3 and S1.Moreover, the gene expression of each subgroup, obtained from the DEG analysis was exhibited in the ideogram of chromosomes using the karyoploteR package (Figure 3C).Enrichment analysis showed that DEGs were significantly enriched in biological processes related to cancer progression.Based on GO analysis, the main biological processes involving the DEGs included ion homeostasis, inorganic cation transmembrane transport, and the regulation of hormone levels.In terms of cellular components, the DEGs were mostly enriched in the external encapsulating structure and extracellular matrix (ECM).In terms of molecular functions, the DEGs were linked by cation transmembrane transport activity, receptor regulator activity, signaling receptor activator activity, etc. (Figures 4A and S2-S6).To further explore the prognostic value of emerging DEGs, we performed univariate Cox proportional hazards regression (Table S2).
Machine Learning Analysis
The results of the ML analysis are shown in Table 1.The deep learning method achieved an accuracy of 97.14%, 97%, 98%, and 92% for predicting CRC in the MSI-H, MSS, chemotherapy, and targeted therapy subgroups, respectively, with AUC values of GSEA analysis showed that there was a relationship between identified DEGs and cell cycle, cell cycle checkpoint, DNA repair, mitotic nuclear division, cellular response to DNA damage stimulus, programmed cell death, epithelial cell differentiation, DNAbinding transcription factor activity, regulation of transcription by RNA polymerase II, Wnt signaling pathway, keratin filaments.According to the Reactom database analysis, DEGs were involved in GPCR signaling and its downstream signaling pathways, the regulation of Insulin-like growth factor (IGF), SLC-mediated transmembrane transport, the degradation of the extracellular matrix (ECM), collagen degradation, biological oxidation, and the activation of matrix metalloproteinases.(Figure 4B,C).
To further explore the prognostic value of emerging DEGs, we performed univariate Cox proportional hazards regression (Table S2).
Machine Learning Analysis
The results of the ML analysis are shown in Table 1.The deep learning method achieved an accuracy of 97.14%, 97%, 98%, and 92% for predicting CRC in the MSI-H, MSS, chemotherapy, and targeted therapy subgroups, respectively, with AUC values of 1.0, 1.0, 1.0, and 0.88.This model had the best performance in the MSI-H and MSS subgroups.Then, 14 candidate genes were identified as novel genes which were dysregulated in both DNA and RNA levels.Also, the candidate genes and common genes resulting from the survival analysis were then displayed on a Venn diagram (Figure 4D and Table S3).Following the visualization described in the MAF data analysis stage, 232 variants from 14 candidate genes related to survival were analyzed (Figure 5).Then, we confirmed the candidate genes in an additional cohort of our patients, which was detected by whole genome sequencing (WES) in 15 cases.Then, 11 genes emerged between the different cohorts, including ASPHD1, C2orf61, C6orf223, CADPS, CCDC150, DCAF4L1, MIA, NEK5, ONECUT3, PNPLA3, and TMEM145 (Table S4).1.0, 1.0, 1.0, and 0.88.This model had the best performance in the MSI-H and MSS subgroups.Then, 14 candidate genes were identified as novel genes which were dysregulated in both DNA and RNA levels.Also, the candidate genes and common genes resulting from the survival analysis were then displayed on a Venn diagram (Figure 4D and Table S3).Following the visualization described in the MAF data analysis stage, 232 variants from 14 candidate genes related to survival were analyzed (Figure 5).Then, we confirmed the candidate genes in an additional cohort of our patients, which was detected by whole genome sequencing (WES) in 15 cases.Then, 11 genes emerged between the different cohorts, including ASPHD1, C2orf61, C6orf223, CADPS, CCDC150, DCAF4L1, MIA, NEK5, ONECUT3, PNPLA3, and TMEM145 (Table S4).
The Prognostic Value of ZBTB12 and ASPHD1
Of note, RNA-seq data certified the dysregulation of candidate genes identified from Ml and DNA-seq and shortlisted ZBTB12 and ASPHD1 as the disease-associated genes (Figure 6).According to the Human Protein Reference Database, ZBTB12 and ASPHD1 interact with HRAS, Ras-associated protein 1, and HRAS, PRRC2A, MSL3, and PIK3CA (Figure 6A,B).The results of WES found nine genetic variants in ASPHD1 and ZBTB1 (Figure 6C,D).According to the RegulomeDB database and 3DSNP, the rs925939730 variant of the ASPHD1 and rs1428982750 variant of the ZBTB1 regulate gene expression and affect chromatin state in the colon and rectum (Tables S5 and S6).Moreover, the rs1428982750 variant was linked to VARS and EHMT2 genes, and the rs925939730 variant was associated with
The Prognostic Value of ZBTB12 and ASPHD1
Of note, RNA-seq data certified the dysregulation of candidate genes identified from Ml and DNA-seq and shortlisted ZBTB12 and ASPHD1 as the disease-associated genes (Figure 6).According to the Human Protein Reference Database, ZBTB12 and ASPHD1 interact with HRAS, Ras-associated protein 1, and HRAS, PRRC2A, MSL3, and PIK3CA (Figure 6A,B).The results of WES found nine genetic variants in ASPHD1 and ZBTB1 (Figure 6C,D).According to the RegulomeDB database and 3DSNP, the rs925939730 variant of the ASPHD1 and rs1428982750 variant of the ZBTB1 regulate gene expression and affect chromatin state in the colon and rectum (Tables S5 and S6).Moreover, the rs1428982750 variant was linked to VARS and EHMT2 genes, and the rs925939730 variant was associated with the MAZ gene (Tables S7 and S8).The rs1428982750 variant of the ZBTB12 gene had a score of 0.60906 for its role in gene expression regulation.Also, this variant affected the state of the chromatin transcription activity in the colon and rectum.Chromatin immunoprecipitation coupled with sequencing (CHIP-seq) results showed that the ZBTB12 gene variant affects the binding site of transcription factors and various regulatory factors.(Figure S7C).The rs925939730 variant of the ASPHD1 gene had a score of 0.77967 for its role in regulating gene expression.Also, this variant affected the state of the chromatin transcription activity in the colon and rectum.CHIP-seq results showed that the ASPHD1 gene variant affects the binding site of transcription factors and various regulatory factors.(Figure 6E).The results of the rs1428982750 variant of the ZBTB12 gene in the 3DSNP database showed that the association of this variant with the regulatory factors of gene expression has a score of 58.4 (Figure S7A).The different positions of this variant.The results of the rs925939730 variant of the ASPHD1 gene in the 3DSNP database showed that the association of this variant with the regulatory factors of gene expression has a score of 59.7 (Figure S7B).
Discussion
Colorectal cancer ranks as the third most common cause of cancer-related mortality [13].Early diagnosis of this disease leads to more effective treatment, reduced treatment costs, reduced disease progression, and decreased morbidity and mortality.Since cancer is intimately linked to genetic alterations, pinpointing these changes is especially critical ROC curve data was obtained by plotting the rate of sensitivity versus specificity.Also, Kaplan-Meier revealed that the overall survival of patients with cancer having low ASPHD1 expression had higher overall survival (OS) than patients with cancer with high ASPHD1 expression (p < 0.05).Similarly, cancers with high ZBTB12 expression were associated with poor patient survival compared to cancers with low ZBTB12 expression (p < 0.05) (Figure 6F,G).As shown in Figure 6H and Tables 2 and 3, ASPHD1, ZBTB12, and their combination were able to discriminate CRC with an area under the curve (AUC) of 0.948, 0.96, and 0.986, respectively.At the cutoff values of 0.863, 0.891, and 0.886, the sensitivities of ASPHD1, ZBTB12, and their combination were 0.878%, 0.861%, and 0.934%, respectively, with specificities of 1.The combination of ASPHD1 and ZBTB12 showed higher AUC and sensitivity than each of these candidate genes alone.To further verify their values, the expression of these two candidate genes was evaluated in an additional cohort of CRC via qRT-PCR.The data showed a significantly higher expression of ASPHD1 and ZBTB12 in CRC tissues (p < 0.05) (Figure 6I).
Discussion
Colorectal cancer ranks as the third most common cause of cancer-related mortality [13].Early diagnosis of this disease leads to more effective treatment, reduced treatment costs, reduced disease progression, and decreased morbidity and mortality.Since cancer is intimately linked to genetic alterations, pinpointing these changes is especially critical for early diagnosis.Implementing the right analyses of gene expression information can promote optimal treatment selection in the early stages of the development of various cancers.Identifying prognostic biomarkers and achieving diagnosis constitute a worthwhile tactic for disease management and care [14,15].Artificial intelligence (AI) and deep learning (DL) are being widely adopted in medicine to enhance diagnosis, treatment, and research on diagnosing colorectal cancer (CRC) has followed this trend.DL is now integrated across CRC diagnostic approaches such as histopathology, endoscopy, radiology, and biochemical blood tests.By automating complex data analysis, DL allows for more precise CRC detection and characterization.Although AI adoption faces regulatory hurdles, it has the potential to optimize the diagnosis of CRC recurrence and personalized care by synthesizing diverse medical data and uncovering new insights.Overall, AI and DL are transforming the management of patients with CRC through improved diagnostic accuracy [16].
Our previous studies identified prognostic and diagnostic biomarkers in colorectal cancer and gastric cancer using RNA-seq analysis and machine learning [17][18][19].In contrast to our previous study, the current study was designed based on an integrated two omics and deep learning approach to identify prognostic and diagnostic biomarkers in colorectal cancer (CRC) patients at different disease stages (early and metastatic).By combining multi-omics data and advanced computational methods, the present study provides novel insights into stratifying CRC patients based on genetic and expression profiles correlated with disease progression and outcomes.To the best of our knowledge, this is the first study showing the potential association of two genetic variants, rs1428982750 in ZBTB12 and rs925939730 in ASPHD1 genes, and the prognostic value of these genes in colorectal cancer.Bian Wu et al. used WES and RNA-seq to indicate prognosis prediction in patients with stage IV colorectal cancer.The results showed the following mutations in the genes: APC, TP53, KRAS, TTN, SYNE1, SMAD4, PIK3CA, RYR2.BRAF did not reveal any significant associations between the mutational status of those genes and patient prognosis [20].Our study revealed that mutations in the genes ZBTB12 and ASPHD1 may serve as potential prognostic markers in patients.Specifically, we demonstrated that the mutational status of ZBTB12 and ASPHD1 was associated with clinical outcomes in the patient cohort examined.Chen et al. analyzed gene expression data from the GEO and TCGA databases and identified 10 hub genes with high diagnostic values based on ROC curve analysis.A ninegene prognostic signature was also identified and shown to predict overall survival [21].Importantly, we validated the expression of ASPHD1 and ZBTB12 genes through qPCR and their variants using whole exome sequencing in additional patient cohorts.
Data from the PPI network showed that ASPHD1 is related to several proteins and genes such as KIF22, INO80E, SEZ6L2, and DOC2A, most of which are cancer-related.Kinesin family member 22 (KIF22) is a regulator of cell mitosis and cellular vesicle transport.It is involved in spindle formation and the movement of chromosomes during mitosis.The alteration of KIF22 is associated with several cancers, including CRC.A previous study indicated that KIF22 is upregulated in CRC samples and that KIF22 expression is correlated with tumors and the clinical stage of CRC.Moreover, the suppression of KIF22 inhibited cell proliferation and xenograft tumor growth [22].
SEZ6L2 regulates cell fate by involving the transcription of type 1 transmembrane proteins.A study showed that SEZ6L2 was significantly upregulated in CRC tissues, and this upregulation was associated with poor prognosis in patients with CRC [23].Lastly, INO80E is involved in transcriptional regulation, DNA replication, and probably DNA repair.Therefore, we hypothesize that ASPHD1 may play a critical role in the pathogenesis of CRC.
PRRT2 is also related to several kinds of human solid tumors [24].The results of the Protein-protein interaction network demonstrated that ZBTB12 is linked to numerous genes, including HRAS, PIK3CA, MSL3, and PRRC2A.
Phosphatidylinositol-4,5-bisphosphate 3-kinase (PI3K), an important kinase involved in the PI3K/AKT1/MTOR pathway, plays a crucial role in the growth and proliferation of various solid tumors, and PIK3CA is one of the most frequently mutated genes in CRC [25].Harvey rat sarcoma viral oncogene homolog (HRAS) is involved in the activation of Ras protein signal transduction, and its mutations can be found in bladder and head and neck squamous cell carcinomas [26].It has been shown that proline-rich coiled-coil2A (PRRC2A) takes part in tumorigenesis and immunoregulation.Recent studies have revealed that PRRC2A impacts pre-mRNA splicing and translation initiation [27].In this context, several studies have demonstrated that there is a relationship between PRRC2A and several kinds of human cancers, such as hepatocellular carcinoma [28] and non-Hodgkin lymphoma [29].
Collectively, ASPHD1 and ZBTB12 are linked to multiple proteins and genes which are associated with cancer initiation and progression.Moreover, our results from WES analysis indicated that the rs925939730 variant of the ASPHD1 gene and the rs1428982750 variant of the ZBTB1 gene regulate gene expression and affect the chromatin state in the colon and rectum.
In addition, our findings demonstrated that there was an interaction between the rs1428982750 variant and VARS and EHMT2 genes.Valyl-tRNA synthetase (VARS) was linked with CRC [30], breast cancer [31], and leukemia [30].Euchromatic histone-lysine N-methyltransferase 2 (EHMT2) methylates histone H3 lysine 9 to generate heterochromatin and inhibit tumor suppressor genes [32].Furthermore, the rs925939730 variant was associated with the MAZ gene.MAZ acts as a transcription factor that can be combined with c-MYC and GA box to regulate the initiation and termination of transcription.The deregulated expression of MYC-associated zinc finger protein (MAZ) is correlated with the progression of tumors such as colorectal adenocarcinoma [33], hepatocellular carcinoma [34], renal cell carcinoma [35], glioblastoma [36], breast carcinoma [37], and prostate adenocarcinoma [38].Altogether, the rs925939730 and rs1428982750 gene variants of ASPHD1 might be involved in gene expression and epigenetic regulation.
Conclusions
Our data show the prognostic value of ASPHD1 and ZBTB12 in CRC, warranting further investigations to validate their clinical potential as prognostic markers and predictive markers for colorectal cancer.Our study had some limitations and challenges, including the difficulty we experienced obtaining access to more patients for evaluating gene expression, carrying out functional studies, and analyzing other omics data to assess important pathways and biological processes in cancer.Expanding our omics approaches beyond just transcriptomics to also include proteomics, metabolomics, etc., would provide a more comprehensive understanding of the key mechanisms in cancer.Overcoming these limitations will be critical for future efforts to elucidate the complex molecular landscape of cancer and identify novel therapeutic targets or biomarkers.
Supplementary Materials:
The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/cancers15174300/s1, Figure S1: PCA plots, volcano plots, and heat maps for DEGs in each subgroup of CRC patients from the TCGA database.S5: The results of the effect of the rs1428982750 variant of the ZBTB12 gene on chromatin status in the colon and rectum.Table S6: The results of the effect of the rs925939730 variant of the ASPHD1 gene on chromatin status in the colon and rectum.Table S7: Results of rs1428982750 variant of ZBTB12 gene in 3DSNP database.Table S8: Results of rs925939730 variant of ASPHD1 gene in 3DSNP database.
Figure 1 .
Figure 1.Visualization and summary of the analysis results of MAF data in the early-stage group (I, II stages) with the maftools package.(A) Bar and box plots display the frequency of different variants across samples (DEL: Deletion, INS: Insertion, SNP: Single-nucleotide polymorphism, ONP: Oligo-nucleotide polymorphism).(B) Oncoplots (note: variants annotated as Multi_Hit are genes that are mutated repeatedly within the same sample).(C) Transition and Transversion mutations (Ti: Transition; Tv: Transversions).(D) A boxplot of Variant Allele Frequencies.(E) Somatic Interactions show results of exclusive/co-occurrence event analysis.(F) Drug-gene interaction analysis based on the Drug-Gene Interaction database.(G) Oncogenic Signaling Pathways.
Figure 1 .
Figure 1.Visualization and summary of the analysis results of MAF data in the early-stage group (I, II stages) with the maftools package.(A) Bar and box plots display the frequency of different variants across samples (DEL: Deletion, INS: Insertion, SNP: Single-nucleotide polymorphism, ONP: Oligo-nucleotide polymorphism).(B) Oncoplots (note: variants annotated as Multi_Hit are genes that are mutated repeatedly within the same sample).(C) Transition and Transversion mutations (Ti: Transition; Tv: Transversions).(D) A boxplot of Variant Allele Frequencies.(E) Somatic Interactions show results of exclusive/co-occurrence event analysis.(F) Drug-gene interaction analysis based on the Drug-Gene Interaction database.(G) Oncogenic Signaling Pathways.Cancers 2023, 15, x FOR PEER REVIEW 8 of 19
Figure 2 .
Figure 2. Visualization and summary of the analysis results of MAF data in the advanced-stage group (IV stage) with the maftools package.(A) Bar and box plots display the frequency of different variants across samples (DEL: Deletion, INS: Insertion, SNP: Single-nucleotide polymorphism, ONP: Oligo-nucleotide polymorphism).(B) Oncoplots (note: variants annotated as Multi_Hit are genes that are mutated repeatedly within the same sample).(C) Transition and Transversion mutations (Ti: Transition; Tv: Transversions).(D) Boxplot of Variant Allele Frequencies.(E) Somatic Interactions show the results of exclusive/co-occurrence event analysis.(F) Drug-gene interaction analysis based on the Drug-Gene Interaction database.(G) Oncogenic Signaling Pathways.
Figure 2 .
Figure 2. Visualization and summary of the analysis results of MAF data in the advanced-stage group (IV stage) with the maftools package.(A) Bar and box plots display the frequency of different variants across samples (DEL: Deletion, INS: Insertion, SNP: Single-nucleotide polymorphism, ONP: Oligo-nucleotide polymorphism).(B) Oncoplots (note: variants annotated as Multi_Hit are genes that are mutated repeatedly within the same sample).(C) Transition and Transversion mutations (Ti: Transition; Tv: Transversions).(D) Boxplot of Variant Allele Frequencies.(E) Somatic Interactions show the results of exclusive/co-occurrence event analysis.(F) Drug-gene interaction analysis based on the Drug-Gene Interaction database.(G) Oncogenic Signaling Pathways.
Figure 3 .
Figure 3.The results of the analysis of differentially expressed genes (DEGs) in colorectal adenocarcinoma (COAD) were generated using R software h ps://www.r-project.org/.(A) The heat map.(B) Principal component analysis (PCA).(C) karyoplot.(D) Volcano plot.GSEA analysis showed that there was a relationship between identified DEGs and cell cycle, cell cycle checkpoint, DNA repair, mitotic nuclear division, cellular response to DNA damage stimulus, programmed cell death, epithelial cell differentiation, DNA-binding transcription factor activity, regulation of transcription by RNA polymerase II, Wnt signaling pathway, keratin filaments.According to the Reactom database analysis, DEGs were involved in GPCR signaling and its downstream signaling pathways, the regulation
Figure 4 .
Figure 4. (A) Gene Ontology (GO), (B) GSEA functional annotation, and (C) Reactome functional pathways in colorectal adenocarcinoma (COAD).The p-value is less than 0.05 and is shown by the color.(D) A Venn diagram indicating the number of survival-related genes and the overlap between the different subgroups.
Figure 4 .
Figure 4. (A) Gene Ontology (GO), (B) GSEA functional annotation, and (C) Reactome functional pathways in colorectal adenocarcinoma (COAD).The p-value is less than 0.05 and is shown by the color.(D) A Venn diagram indicating the number of survival-related genes and the overlap between the different subgroups.
Figure 5 .
Figure 5. (A) A Venn diagram shows the count of variants for 14 candidate genes which are common between DNA-seq and RNA-seq analysis.(B) Bar and box plots displaying the frequency of different variants across samples.(C) Oncoplots.(D) Transition and Transversion mutations.(E) Boxplot of Variant Allele Frequencies.(F) Somatic Interactions show the results of exclusive/co-occurrence event analysis.(G) Drug-gene interaction analysis based on the Drug-Gene Interaction database.
Figure 5 .
Figure 5. (A) A Venn diagram shows the count of variants for 14 candidate genes which are common between DNA-seq and RNA-seq analysis.(B) Bar and box plots displaying the frequency of different variants across samples.(C) Oncoplots.(D) Transition and Transversion mutations.(E) Boxplot of Variant Allele Frequencies.(F) Somatic Interactions show the results of exclusive/co-occurrence event analysis.(G) Drug-gene interaction analysis based on the Drug-Gene Interaction database.
19 Figure 6 .
Figure 6.(A,B) Protein-protein interaction (PPI) network of the two genes (ZBTB12, ASPHD1) identified by survival analysis from STRING.(C,D) The different types of ZBTB12 and ASPHD1 variants, along with their respective alterations in the amino acid sequence on chromosomes, as well as the rate of somatic mutation.(E) CHIP-seq results have shown that variants of two genes (ZBTB12, AS-PHD1) affect the binding site of transcription factors and various regulatory factors from the Regu-
Figure 6 .
Figure 6.(A,B) Protein-protein interaction (PPI) network of the two genes (ZBTB12, ASPHD1) identified by survival analysis from STRING.(C,D) The different types of ZBTB12 and ASPHD1 variants, along with their respective alterations in the amino acid sequence on chromosomes, as well as the rate of somatic mutation.(E) CHIP-seq results have shown that variants of two genes (ZBTB12, AS-PHD1) affect the binding site of transcription factors and various regulatory factors from the Regulome DB Database.(F,G) Kaplan-Meier plot of ZBTB12 and ASPHD1 with a prognostic value, pvalue < 0.05.(H) ROC curve analysis revealed the biomarker potency of ZBTB12 and ASPHD1 individually and together using R 4.3.1'scombioROC package.(I) qRT-PCR results indicate that the expression levels of the two genes (ZBTB12 and ASPHD1) are elevated in tumor tissue compared to non-neoplastic tissue.***p > 0.01; **** p > 0.001.
Figure 6 .
Figure 6.(A,B) Protein-protein interaction (PPI) network of the two genes (ZBTB12, ASPHD1) identified by survival analysis from STRING.(C,D) The different types of ZBTB12 and ASPHD1 variants, along with their respective alterations in the amino acid sequence on chromosomes, as well as the rate of somatic mutation.(E) CHIP-seq results have shown that variants of two genes (ZBTB12, ASPHD1) affect the binding site of transcription factors and various regulatory factors from the Regulome DB Database.(F,G) Kaplan-Meier plot of ZBTB12 and ASPHD1 with a prognostic value, p-value < 0.05.(H) ROC curve analysis revealed the biomarker potency of ZBTB12 and ASPHD1 individually and together using R 4.3.1'scombioROC package.(I) qRT-PCR results indicate that the expression levels of the two genes (ZBTB12 and ASPHD1) are elevated in tumor tissue compared to non-neoplastic tissue.*** p > 0.01; **** p > 0.001.
Figure S2: Pathway enrichment analyses of DEGs in MSI-L CRC patients from the TCGA database.
Figure S3: Pathway enrichment analyses of DEGs in MSI-H CRC patients from the TCGA database.
Figure S4: Pathway enrichment analyses of DEGs in MSS CRC patients from the TCGA database.
Figure S5: Pathway enrichment analyses of DEGs in Receiving targeted therapies CRC patients from the TCGA database.
Figure S6 :
Pathway enrichment analyses of DEGs in CRC patients receiving chemotherapy from the TCGA database.
Figure S7: Results of the RegulomeDB and 3DSNP database.
Table 1 .
Results of machine learning analysis.
Table 1 .
Results of machine learning analysis.
Table 2 .
The area under the curve (AUC) and a cut-off value of ASPHD1, ZBTB12, and their combination in CRC.
Table 3 .
Results for the ROC curve for ASPHD1, ZBTB12, and their combination in CRC.
Table S1 :
Number of DEGs in patients with CRC.Table S2: Number of survival-related DEGs.Table S3: Candidate genes based on database search.Table S4: Variants of candidate genes in the Whole Exome Sequencing (WES) data of 15 patients with CRC in the Mashhad population.Table | 9,077.6 | 2023-08-28T00:00:00.000 | [
"Biology"
] |
The Authors Matter: Understanding and Mitigating Implicit Bias in Deep Text Classification
It is evident that deep text classification models trained on human data could be biased. In particular, they produce biased outcomes for texts that explicitly include identity terms of certain demographic groups. We refer to this type of bias as explicit bias, which has been extensively studied. However, deep text classification models can also produce biased outcomes for texts written by authors of certain demographic groups. We refer to such bias as implicit bias of which we still have a rather limited understanding. In this paper, we first demonstrate that implicit bias exists in different text classification tasks for different demographic groups. Then, we build a learning-based interpretation method to deepen our knowledge of implicit bias. Specifically, we verify that classifiers learn to make predictions based on language features that are related to the demographic attributes of the authors. Next, we propose a framework Debiased-TC to train deep text classifiers to make predictions on the right features and consequently mitigate implicit bias. We conduct extensive experiments on three real-world datasets. The results show that the text classification models trained under our proposed framework outperform traditional models significantly in terms of fairness, and also slightly in terms of classification performance.
Introduction
Many recent studies have suggested that machine learning algorithms can learn social prejudices from data produced by humans, and thereby show systemic bias in performance towards specific demographic groups or individuals (Mehrabi et al., 2019;Blodgett et al., 2020;Shah et al., 2020). As one machine learning application, text classification has been proven to be discriminatory towards certain groups of people (Dixon et al., 2018;Borkan et al., 2019). Text classification applications such as sentiment analysis and hate speech detection are common and widely used in our daily lives. If a biased hate speech detection model is deployed by a social media service provider to filter users' comments, the comments related to different demographic groups can have uneven chances to be recognized and removed. Such a case will cause unfairness and bring in negative experience to its users. Thus, it is highly desired to mitigate the bias in text classification.
The majority of existing studies on bias and fairness in text classification have mainly focused on the bias towards the individuals mentioned in the text content. For example, in (Dixon et al., 2018;Park et al., 2018;Zhang et al., 2020), it is investigated how text classification models perform unfairly on texts containing demographic identity terms such as "gay" and "muslim". In such scenarios, the demographic attributes of the individuals subject to bias explicitly exist in the text. In this work, we refer to this kind of bias as explicit bias. Bias in texts, however, can be reflected more subtly and insidiously. While a text may not contain any reference to a specific group or individual, the content can somehow be revealing of the demographic information of the author. As shown in (Coulmas, 2013;Preoţiuc-Pietro and Ungar, 2018), the language style (e.g., wordings and tone) of a text can be highly correlated with its author's demographic attributes (e.g., age, gender, and race). We find that a text classifier can learn to associate the content with demographic information and consequently make unfair decisions towards certain groups. We refer to such bias as implicit bias. Table 1 demonstrates an example of implicit bias. There are two short texts where the first text is written by a white American and the second one by an African American. The task is to predict the sentiment of a text by a convolutional neural network (CNN) model. Words with a red background indicate those with the salient predictive capability by the model where the darker the color, the more salient the words. The words "yup" and "goin" in the second text are commonly used by African Americans (Liu et al., 2020a) and are irrelevant to the sentiment. However, the CNN model has hinted at them and consequently has predicted a positive text to be negative.
In this work, we aim to understand and mitigate implicit bias in deep text classification models.
One key source of bias is the imbalance of training data (Dixon et al., 2018;Park et al., 2018). Thus, existing debiasing methods mainly focus on balancing the training data, such as adding new training data (Dixon et al., 2018) and augmenting data based on identity-term swap (Park et al., 2018). However, these methods cannot be directly applied to mitigate implicit bias. Obtaining new texts from authors of various demographic groups is very expensive. It requires heavy human labor. Meanwhile, given that there is no explicit demographic information in texts, identity-term swap data augmentation is not applicable. Thus, we propose to enhance deep text classification models to mitigate implicit bias in the training process. To achieve this goal, we face tremendous challenges. First, to mitigate the implicit bias, we have to understand how deep models behave. For example, how they correlate implicit features in text with demographic attributes and how the models make biased predictions. Second, we need to design new mechanisms to take advantage of our understandings to mitigate the implicit bias in deep text classifiers.
To address the above challenges, in this paper, we first propose an interpretation method, which sheds light on the formation mechanism of implicit bias in deep text classification models. We show that the implicit bias is caused by the fact that the models make predictions based on incorrect language features in texts. Second, based on this finding, we propose a novel framework Debiased-TC (Debiased Text Classification) to mitigate the implicit bias of deep text classifiers. More specifically, we equip the deep classifiers with an additional saliency selection layer that first determines the correct language features which the model should base on to make predictions. We also propose an optimization method to train the classifiers with the saliency selection layer. Note that both our proposed interpretation method and the learning framework are model-agnostic, where they can be applied to any deep text classifier. We evaluate the framework with two popular deep text classification models across various text classification tasks on three public datasets. The experimental results demonstrate that our method significantly mitigates the implicit bias while maintaining or even improving their prediction performance.
Preliminary Study
In this section, we perform a preliminary study to validate the existence of implicit bias in deep text classification models. We first introduce the data and text classification tasks, and then present the empirical results.
Data and Tasks
In this preliminary study, we investigate different text classification tasks and various demographic groups to validate the implicit bias. We use three datasets, including the DIAL and PAN16 datasets processed by (Elazar and Goldberg, 2018) and the Multilingual Twitter Corpus (MTC) introduced in (Huang et al., 2020).
The DIAL dataset contains dialectal texts collected from Twitter. Each tweet's text is associated with the race of the author as the demographic attribute, denoted as "white" and "black", respectively. This dataset is annotated for two classification tasks: sentiment analysis and mention detection. The sentiment analysis task aims to categorize a text as "happy" or "sad". The mention detection task tries to determine whether a tweet mentions another user, which can also be viewed as distinguishing conversational tweets from non-conversational ones.
The PAN16 dataset consists of tweets. For each tweet, age and gender of its author have been manually labelled. The demographic attribute age has two categories of (18-34) and (≥ 35), and gender has male and female. Also, this dataset is annotated for the mention detection task as described above.
The MTC dataset contains multilingual tweets for the hate speech detection task. Each tweet is annotated as "hate speech" or "non hate speech" and associated with four author's demographic attributes: race, gender, age, and country. We only use the English corpus with the attribute race. In this dataset, the attribute race has two categories, i.e., white and nonwhite.
More statistical information on these three datasets and the links to downloadable versions of the data can be found in Appendix A.
Empirical study
In this subsection, we aim to empirically study if text classification models make the predictions dependent on the demographic attributes of the authors of the texts. The explicit bias in text classification tasks stems from the imbalance of training data (Dixon et al., 2018;Park et al., 2018). For example, when there are more negative examples from one group in the training data, the model learns to correlate that group with the negative label, which results in bias. Inspired by this observation, to validate the existence of implicit bias, we investigate if the imbalance of training data in terms of demographic attributes of the authors can lead to biased predictions. To answer this question, we consider the following setting: (1) the training data has an equal number of positive and negative examples; and (2) positive and negative examples in the training data are imbalanced among different groups of the authors according to their demographic attributes. Intuitively, if the predictions are independent of the demographic attributes of authors, the model should still perform similarly for different groups.
For each task and demographic attribute of authors, we consider two labels (i.e., positive and negative) and two demographic groups (i.e., Group I and Group II). For each dataset, we follow the aforementioned setting to build a training set. We make the training set overall balanced in terms of the labels and demographic groups. That is, we set the overall ratio of positive and negative examples as 1:1, and the overall ratio of examples from Group I and Group II as 1:1 as well. Meanwhile, we make the data in each group imbalanced. In particular, for Group I, we set the ratio of its positive and negative examples to 4:1, while the ratio is automatically set to 1:4 for Group II. We name the proportion of positive and negative samples in Group I as "balance rate". We train a convolution neural network (CNN) text classifier as a representative model on the training set and evaluate it on the test set. We use the false positive/negative rates (Dixon et al., 2018) and the demographic parity rate (a.k.a., positive outcome rate, the probability of the model predicting a positive outcome for one group) (Dwork et al., 2012;Kusner et al., 2017) to evaluate the fairness of the classification models.
The results are shown in Table 2. For the demographic attribute race, Group I/Group II stands for white/black in the DIAL dataset, and white/nonwhite in the MTC dataset. For gender and age, Group I/Group II stands for male/female and age ranges (18-34)/(≥35), respectively. From the table, we observe that in terms of different tasks and demographic attributes of authors, the model shows significant bias with the same pattern. For all cases, the demographic group with more positive examples (Group I) always gets a higher false positive rate, a lower false negative rate, and a higher demographic parity rate than the other group. This demonstrates that imbalanced data can cause implicit bias, and the predictions are not independent of the demographic attributes of authors. Since the text itself doesn't explicitly contain any demographic information, the model could learn to recognize the demographic attributes of authors based on implicit features such as language styles and associate them with a biased outcome. Next, we will understand one formation of implicit bias and then propose Debiased-TC to mitigate it.
Understanding Implicit Bias
In this section, we aim to understand the possible underlying formation mechanism of the implicit bias. Our intuition is -when a training set for sentiment analysis has more positive examples from white authors and more negative examples from black authors, a classification model trained on such a dataset may learn a "shortcut" (Mahabadi et al., 2020) to indiscriminately associates the language style features of white people with the positive sentiment and those of black people with the negative sentiment. In other words, the model does not use the correct language features (e.g., emotional words) to make the prediction. Thus, we attempt to examine the following hypothesis: A deep text classification model presents implicit bias since it makes predictions based on language features that should be irrelevant to the classification task but are correlated with a certain demographic group of authors. To verify this hypothesis, we first propose an interpretation method to detect the salient words a text classification model relies on to make the prediction. The interpretation model enables us to check the overlapping between the salient words and the words related to the authors' demographic attributes. Consequently, it allows us to understand the relationship between such overlapping and the model's implicit bias.
An Interpretation Method
We follow the idea of the learning-based interpretation method L2X (Chen et al., 2018) to train an explainer to interpret a given model. The reasons for choosing L2X are -1) as a learning-based explainer, it learns to globally explain the behavior of a model, instead of explaining a single instance at one time; and 2) the explainer has the potential to be integrated into our debiasing framework to mitigate implicit bias in an end-to-end manner, which will be introduced in Section 4.
A binary text classification model M : . For a certain model M, we seek to specify the contribution of each word in X for M to make the prediction Y . The contributions can be denoted as a saliency distribution S = (s 1 , s 2 , . . . , s n ), where s i is the saliency score of the word x i , and n i=1 s i = 1. Given a model M, we train an explainer E M : X → S to estimate the saliency distribution S of an input text X.
The explainer is trained by maximizing I(X S , Y ), the mutual information (Cover, 1999) between the response variable Y and the selected feature X S of X under saliency distribution S. The selected feature X S = X S = (s 1 · x 1 , s 2 · x 2 , . . . , s n · x n ) 1 is calculated as the element-wise product between X and S. In our implementation, we parametrize the explainer by a bi-directional RNN followed by a linear layer and a Softmax layer. More details about the optimization of the explainer can be found in Appendix B.
Saliency Correlation Measurement
In this work, we assume that the text classification task is totally independent of the demographic attribute of the author of the text. In other words, language features that reflect the author's demographic information should not be taken as evidence for the main task. Thus, we propose to understand the implicit bias of a deep text classification model by examining the overlapping between salient words for the main task and the words correlated with the 1 Without confusion, we use xi to denote both a word and its word embedding vector. With the interpretation model, we can estimate the saliency distributions of the input words for the classification task and the demographic attribute prediction task, respectively, and then check their overlapping. As shown in Figure 1, we train two models M Y and M Z with the same architecture for the former and the latter tasks, respectively. Then, two corresponding explainers E Y and E Z are trained for them. Thus, given an input text X, two explainers can estimate the saliency distributions S Y and S Z on two tasks, respectively. We use the Jensen-Shannon (JS) divergence JS(S Y ||S Z ) to measure the overlap between language features that these two tasks relying on to make the predictions on Y and Z.
Empirical Analysis
In this subsection, we present the experiments to verify our hypothesis on the formulation of implicit bias. Following the experimental settings in Section 2.2, we vary the "balance rate" of the training data and then observe how the saliency correlation changes. We use CNN text classifiers (see 5.2 for details) for both M Y and M Z . In Figure 2, we show how the average JS divergence and the demographic parity difference (DPD) vary with the changes of the balance rate. DPD is the absolute value of the difference between the demographic parity rates for the two groups. We only report the results for DIAL and PAN16 datasets and DPD as the fairness metric since we achieved similar results for other settings. For each task and each demographic attribute, the DPD is small when the training data are balanced and becomes large when the data are imbalanced. However, the JS divergence is large for balanced data while small for imbalanced data. A larger DPD indicates stronger implicit bias and a smaller JS divergence stands for a stronger overlap between the saliency distributions for the two tasks. Thus, these observations suggest that when the training data are imbalanced, the text classifiers tend to use language features related to the demographic attribute of authors to make the prediction.
The Bias Mitigation Framework
In the previous section, we showed that a model with implicit bias tends to utilize features related to the demographic attribute of authors to make the prediction, especially when training data is imbalanced in terms of the demographic attribute of authors. One potential solution is to balance the training data by augmenting more examples from underrepresented groups. However, collecting new data from authors of different demographics is expensive. Thus, to mitigate the implicit bias, we propose the novel framework Debiased-TC. Our proposed approach can mitigate implicit bias by automatically correcting their selection of input features. In this section, we will first introduce the proposed framework with the corresponding optimization method.
Debiased Text Classification Model
An illustration of Debiased-TC is shown in Figure 3. Similar to the explainer in the interpretation model, we equip the base model Figure 3: An illustration of the bias mitigation model.
a corrector layer C after the input layer. The corrector C : X → S learns to correct the model's feature selection. It first maps an input text X = (x 1 , x 2 , . . . , x n ) to a saliency distribution S = (s 1 , s 2 , . . . , s n ), which is expected to give high scores to words related to the main tasks and low scores to words related to demographic attributes of authors. Then, it assigns weights to the input features with the saliency scores by calculating X S = X S, which is fed into the classification model M Y for prediction.
To train a corrector to achieve the expected goal, we adopt the idea of adversarial training. More specifically, in addition to the main classifier M Y , we introduce an adversarial classifier M Z , which takes X S as the input and predicts the demographic attribute Z. During the adversarial training, the corrector attempts to help M Y make correct predictions while preventing M Z from predicting demographic attributes. To make this feasible, we use the gradient reversal technique (Ganin and Lempitsky, 2015), where we add a gradient-reversal layer between the weighted inputs X S and the adversarial classifier M Z . The gradient-reversal layer has no effect on its downstream components (i.e., the adversarial classifier M Z ). However, during backpropagation, the gradients that pass down through this layer to its upstream components (i.e., the corrector C) are getting reversed. As a result, the corrector C receives opposite gradients from M Z . The outputs of the M Y and M Z are used as signals to train the corrector such that it can upweight the words correlated with the main task label Y and downweight the words correlated with the demographic attribute Z. We set the adversarial classifier M Z with the same architecture as the main classifier M Y . The corrector C has the same architecture as the explainer introduced in Section 3.
An Optimization Method for Debiased-TC
In this subsection, we discuss the optimization method for the proposed framework. We denote the parameters of M Y , M Z and C as W Y , W Z and Θ, respectively. The optimization task is to jointly optimize the parameters of the classifiers, i.e., W Y and W Z , and the parameters of the corrector, i.e., Θ. We can view the optimization as an architecture search problem. Since our debiasing framework is end-to-end and differentiable, we develop an optimization method for our framework based on the differentiable architecture search (DARTS) techniques (Liu et al., 2018;Zhao et al., 2020). We update M Y , M Z by optimizing the training losses L Y train and L Z train on the training set and update Θ by optimizing the validation loss L val on the validation set through gradient descent. We denote the cross-entropy losses for M Y and M Z as L Y and L Z , respectively. L Y train and L Z train indicate the cross-entropy losses L Y and L Z on the training set. L val denotes the combined loss of the two crossentropy losses L = L Y + L Z on the validation set.
The goal of optimizing the corrector is to find optimal parameters Θ * that minimizes the validation loss L val (W Y * , W Z * , Θ), where the optimal parameters W Y * and W Z * are obtained by minimizing the training losses as follows.
The above goal forms a bi-level optimization problem (Maclaurin et al., 2015;Pham et al., 2018), where Θ is the upper-level variable and W Y and W Z are the lower-level variables: Optimizing Θ is time-consuming due to the expensive inner optimization of W Y and W Z . Therefore, we leverage the approximation scheme as DARTS: where ξ is the learning rate for updating W Y and W Z . The approximation scheme estimates W Y * (Θ) and W Z * (Θ) by updating W Y and W Z for a single training step, which avoids total optimization W * (Θ) = arg min W L train (W, Θ * ) to the convergence. In our implementation, we apply first-order approximation with ξ = 0, which can even lead to more speed-up. Also, in our specific experiments, since the amount of validation data is limited, we build an augmented validation dataset V = V ∪ T combining the original validation set V with the training set T for optimizing Θ. More details of the DARTS-based optimization algorithm are shown in Appendix C.
Experiment
In this section, we conduct experiments to evaluate our proposed debiasing framework. Through the experiments, we try to answer two questions: 1) Does our framework effectively mitigate the implicit bias in various deep text classification models? and 2) Does our framework maintain the performance of the original models (without debasing) while reducing the bias?
Baselines
In our experiments, we compare our proposed debiasing framework with two baselines. Since there is no established method for mitigating implicit bias, we adopt two debiasing methods designed for traditional explicit bias and adapt them for implicit bias.
Data Augmentation* (Data Aug) (Dixon et al., 2018). We manually balance the training data of two demographic groups by adding sufficient negative examples for Group I and positive examples for Group II. As a result, the ratio of positive and negative training examples for both groups is 1:1. As discussed in the introduction, obtaining additional labeled data from specific authors is very expensive. In this work, we seek to develop bias mitigation methodology without extra data. Since Data Aug introduces more training data, it's not fair to directly compare it with other debiasing methods that only utilize original training data (including our method). We include Data Aug as a special baseline for reference.
Instance Weighting (Ins Weigh) (Zhang et al., 2020). We re-weight each training instance with a numerical weight P (Y ) P (Y |Z) based on the label distribution for each demographic group to mitigate explicit bias. In this method, a random forest classifier is built to estimate the conditional distribu-tion P (Y |Z) and the marginal distribution P (Y ) is manually calculated.
Experimental Settings
We conduct our experiments for implicit bias mitigation on two representative base models: CNN (Kim, 2014) and RNN (Chung et al., 2014). We use the same datasets with manually designed proportions, as described in Section 2.2. The details of the base models, as well as the implementation details for the replication of the experiments, can be found in Appendix D.
Performance Comparison
We train the base models with our proposed debiasing framework as well as the baseline debiasing methods. We report the performance on the test set in terms of fairness and classification performance. Fairness Evaluation. Table 3 shows the results for fairness evaluation metrics: false positive equality difference (FPED), false negative equality difference (FNED), and DPD. FPED/FNED indicates the absolute value of the difference between the false positive/negative rates of the two groups. We make the following observations. First, the base models attain high FPED, FNED, and DPD, which indicates the existence of significant implicit bias towards the authors of the texts. Ins Weigh seems ineffective in mitigating implicit bias since it only achieved comparable fairness scores with the base models. Note that not every example that belongs to a certain group necessarily results in bias towards that group. Thus, assigning a uniform weight for all examples with the same label Y and demographic attribute Z is not a proper way to reduce implicit bias. Third, both Data Aug and Debiased-TC can mitigate the implicit bias by achieving lower equality and demographic parity differences. However, compared to Data Aug, Debiased-TC has two advantages. First, Data Aug needs to add more training data while Debiased-TC does not. Debiased-TC can locate the main source of implicit bias by analyzing how it forms in a deep text classification model. Due to the proposed corrector model, it can make a classification model focus on the relevant features for predictions and discard the features that may lead to implicit bias. Second, Debiased-TC is more stable than Data Aug. For the sentiment classification task with race as the demographic attribute, the CNN and RNN classifiers trained on augmented data still result in high FPED and DPD scores. This suggests that balancing the training data cannot always mitigate implicit bias. In fact, only training examples with demographic language features can contribute to the implicit bias. Since some texts in the training set do not contain any language features belonging to a demographic group, they do not help balance the data. Text Classification Performance Evaluation. The prediction performance of the text classification models trained under various debiasing methods is shown in Table 4, where we report the accuracy and F1 scores. First, it is not surprising to see that Data Aug achieves the best performances, since the data augmentation technique introduces more training data. It's not fair to directly compare it with other debiasing methods that only utilize original training data. Second, in most cases, our method achieves comparable or even better performance than the original base models. As we verified before, the implicit bias of a text classification model is caused by the fact that it learns a wrong correlation between labels and demographic language features. Debiased-TC corrects the model's selection of language features for predictions and thereby improves its performance on the classification task.
In conclusion, our proposed debiasing framework significantly mitigates the implicit bias, while maintaining or even slightly improve the classification performance.
Related Work
Fairness in NLP. Recent research has demonstrated that word embeddings exhibit human biases for text data. For example, in word embeddings trained on large-scale real-world text data, the word "man" is mapped to "programmer" while "woman" is mapped to "homemaker" (Bolukbasi et al., 2016). Some works extend the research of biases in word embeddings to that of sentence embeddings. The work (May et al., 2019) examines popular sentence encoding models from CBoW, GPT, ELMo to BERT, and shows that those models inherit human's prejudices from the training data. For the task of coreference resolution, a benchmark named WinoBias is proposed (Zhao et al., 2018) to measure the gender biases with a debiasing method based on data augmentation. Prates et al. (2018) reveal that Google's machine translation system shows gender biases in various languages. Existing debiasing methods for word embeddings are adopted to mitigate the biases in machine translation systems . In the task of dialogue generation, it is first studied by (Liu et al., 2020a) on the biases learned by dialogue agents from human conversation data. It is shown that significant gender and race biases exist in popular dialogue models. As a countermeasure, Liu et al. (2020b) propose to mitigate gender bias in neural dialogue models with adversarial learning.
Fairness in Text Classification. For the text classification problem, Dixon et al. (2018) demonstrate that the source of unintended bias in models is the imbalance of training data, and they provide a debiasing method, which introduces new data to balance the training data. In (Park et al., 2018), gender biases are measured on abusive language detection models, and the effect of different pre-trained word embeddings and model architectures are analyzed. By considering the various ways that a classifier's score distribution can vary across designated groups, a suite of threshold-agnostic metrics is introduced in (Borkan et al., 2019), which provides a nuanced view of this unintended bias. Furthermore, the work (Zhang et al., 2020) proposes to debias text classification models using instance weighting, i.e., different weights are assigned to the training samples involving different demographic groups. The works discussed above focus on explicit bias, where the demographic attributes are explicitly expressed in the text. However, works studying implicit bias are rather limited. Huang et al. (2020) introduce the first multilingual hate speech dataset with inferred author demographic attributes. Through experiments on this dataset, they show that popular text classifiers can learn the bias towards the demographic attribute of the author. But this work doesn't discuss how the bias is produced, and no debiasing method is provided.
Conclusion
In this paper, we demonstrate that a text classifier with implicit bias makes predictions based on language features correlated with demographic groups of authors, and propose a novel learning framework Debiased-TC to mitigate such implicit bias. The experimental results show that Debiased-TC sig-nificantly mitigates implicit bias, and maintains or even improves the text classification performance of the original models. In the future, we will investigate implicit bias in other NLP applications.
B Optimization of the Explainer
We train the explainer E by maximizing the mutual information between the response variable Y and the selected features X S . The optimization problem can be formulated as: Solving the optimization problem in Eq.
(2) is equivalent to finding an explainer E satisfying the following: Hence, we train the explainer E by optimizing P M (Y |X S ) with the parameters of the classification model M fixed. In our implementation, we adopt the cross-entropy loss for training, as we do when we train the classification model M.
C An Optimization Method for Debiased-TC
We present our DARTS-based optimization algorithm in Algorithm 1. In each iteration, we first update the corrector's parameters based on the augmented validation set V (lines 2-3). Then, we collect a new mini-batch of training data (line 4). We generate the saliency scores S = (s 1 , s 2 , . . . , s n ) for the training examples via the corrector with its current parameters (line 5). Next, we make predictions via the classifiers with their current parameters and X S (line 6). Eventually, we update the parameters of the classifiers (line 7). 2 Output: classifier parameters W Y * and W Z * ; and corrector parameters Θ * 3 Initialize W Y , W Z and Θ 1: while not converged do 2: Sample a mini-batch of validation data from V = V ∪ T 3: Update Θ by descending train (W Z , Θ), Θ (ξ = 0 for first-order approximation) 4: Collect a mini-batch of training data from T 5: Generate S via the corrector with current parameters Θ 6: Generate predictions via the classifiers with current parameters W Y , W Z and XS 7: Update W Y and W Z by descending ∇ W Y L Y train (W Y , Θ) and ∇ W Z L Z train (W Z , Θ) 8: end while D Implementation Details
D.1 Details of Base Models
In the base model CNN, we use 100 filters with three different kernel sizes (3, 4, and 5) in the convolution layer, where we use a Rectified Linear Unit (ReLU) as the non-linear activation function. Each obtained feature map is processed by a maxpooling layer. Then, the features are concatenated and fed into a linear prediction layer to get the final predictions. A dropout with a rate of 0.3 is applied before the linear prediction layer.
For the base model RNN, we use a one-layer unidirectional RNN with Gated Recurrent Units (GRU). The hidden size is set to 300. The last hidden state of the RNN is fed into a linear prediction layer to get the final predictions. We apply a dropout with a rate of 0.2 before the linear prediction layer.
D.2 Details of Experimental Settings
For the text classifiers, we use randomly initialized word embeddings with a size of 300. All the models are trained by an Adam optimizer (Kingma and Ba, 2014) with an initial learning rate of 0.001. We apply gradient clipping with a clip-value of 0.25 to prevent the exploding gradient problem. The batch size is set to 64. For the base model and the baseline methods, when the prediction accuracy of the validation data doesn't improve for 5 consecutive epochs, the training is terminated, and we pick the model with the best performance on the validation set. Our model utilizes the validation data for train-ing. To avoid it overfitting the validation data, we don't select the model based on its performance on the validation set. Instead, we train the model for a fixed number of epochs (5 epochs, the same for all the three datasets) and evaluate the obtained model. | 7,999.6 | 2021-05-06T00:00:00.000 | [
"Computer Science"
] |
Preformulation studies for the development of a microemulsion formulation from Ambrosia peruviana All., with anti-inflammatory effect
Abstract Natural products are considered an important source of the therapeutic arsenal currently available. Among these alternatives are the seeds of Ambrosia peruviana (altamisa), whose extract has shown an anti-inflammatory effect. The main objective of this work was to perform a preformulation study of Ambrosia peruviana seeds ethanolic extract, where the main factors that affect the physical, chemical, and pharmacological stability of the extract were evaluated, as well as a compatibility study by differential scanning calorimetry (DSC) analysis against different excipients. A dry extract was obtained by rotary evaporation of the seeds macerated with 96% ethanol. The anti-inflammatory activity was determined by measuring its effect on NO production in RAW 264.7 macrophages, stimulated with LPS. The results showed that the dry extract maintained its stability over time when stored at a temperature of 4 and 25ºC, demonstrating its biological activity, the content of phenolic compounds, and its physicochemical parameters remain practically invariable. However, when exposed to high temperatures (60 ºC) it was affected. The thermal analysis revelated that the behavior of most of the selected excipients and the dry extract was maintained, which indicates that it did not present incompatibilities, therefore they can be candidates for formulating a microemulsion.
INTRODUCTION
Nowadays a large part of the research aimed at the development of new and powerful therapeutic agents is focused on the study of plants used in traditional medicine.In fact, in defining what is considered a drug derived from a natural product, it is estimated that between 25-50% of pharmaceutical products currently used are derived from natural products (Kingston, 2010).Those products have taken hold in the market because the population believes that they have fewer adverse effects than synthetic drugs.The WHO recognizes their great relevance in primary care programs since they represent effective alternatives to prevent or treat several diseases (OMS, 2013).
Colombia is a country with a wide variety of ecosystems, ranking among the fourteen countries with the highest biodiversity index, with a fundamental role in sustaining living beings and also representing a source with excellent research potential, making use of plants for the development of new products, such as Preformulation studies for the development of a microemulsion formulation from Ambrosia peruviana All., with anti-inflammatory effect Page 2/14 Braz.J. Pharm.Sci.2023;59: e22505 Yuri Palacio, Jenny-Paola Castro, Valquiria Linck Bassani, Luis Alberto Franco, Carlos-Alberto Bernal phytopharmaceuticals (Andrade, 2011), especially those that present a high demand, such as phytopharmaceuticals with anti-inflammatory activity.The high demand for compounds with this activity is because many studies carried out in recent decades have established a clear relationship between the inflammatory response and the physiopathology of chronic non-transmissible diseases (Pan, Lai, Ho, 2010;Prasad, Sung, Aggarwa, 2012).The inflammation in these diseases may appear to different degrees and could be associated with different factors such as age, genre, genetic susceptibility, and modifiable risk factors related to lifestyle (Medzhitov, 2008;Meetoo, 2008).The most clinically meaningful drugs for the treatment of these diseases are steroids and non-steroidal anti-inflammatory drugs (NSAIDs), which have a high incidence of side effects.(Nagai et al., 2017;Vonkeman, Van de Laar, 2010).These reasons have encouraged the search for promising, safer, and more effective substances to treat diseases involving inflammation.
Ambrosia peruviana (Asteraceae), is an aromatic plant, used in traditional medicine in several South American countries for the treatment of different diseases that cause inflammation such as: colic, chronic pain, arthritis, spasms, and infections, among others.(Cicció, Chaverri, 2015;Jimenez-Usuga et al., 2016).In vitro studies on this plant species have shown that the ethanolic extract of seeds has showed significant inhibition on inflammatory markers in RAW 264.7 macrophages stimulated with LPS.(Castro, Franco, Diaz, 2021).Those antecedents make it an important starting point for the development of herbal medicine with anti-inflammatory properties.
The development of a phytopharmaceutical product involves objectives aimed at obtaining a physically and chemically stable, technologically feasible, and biologically available product, for which the concept of "Rational Design of Medicines" must be applied, which in its first stage includes a preformulation study.It is a useful tool in the development of conventional medicines and turn have many elements in common with the preformulation of phytopharmaceuticals (Wyttenbach et al., 2005).However, some of these elements are difficult to standardize, due to the complex composition that characterizes these preparations (Bernal, Ramos, Baena, 2019;Matiz, Cárdenas, Rincon, 2007).These studies must be carried out as a stage prior to starting the formulation phase.At this stage, the characterization of the physicochemical properties and the study of the interaction of the extract with possible excipients to be used in the formulation should be performed.At this level, there must be analytical methodologies to be able to ensure the safety, efficacy, and stability of the product in the subsequent stages.In addition, because of the collection of all this information, the guidelines to consider in the implementation of the product manufacturing process in the later stages are obtained (Bernal, Ramos, Baena, 2019;Kopelman, Augsburger, 2002;Wu et al., 2009).Considering the panorama previously described, in this work a preformulation study of the ethanolic extract of Ambrosia peruviana seeds was carried out with focus on obtaining a microemulsion for topical use with antiinflammatory effects.
Plant material
Ambrosia peruviana seeds were collected in Turbana, Bolívar, Colombia (10°24′0″ N, 75°30′0″ O), in July 2018.A voucher specimen of the plant was identified in the herbarium of the Universidad de Antioquia, (Medellin-Colombia) by the biologist Felipe Cardona and deposited with the identification code HUA 214539.
Plant extraction and phytochemical screening
Dried and powdered seeds of A. peruviana were exhaustively extracted with ethanol (96% v/v) by maceration at room temperature (25±3 °C).The extract obtained was dried in a rotary evaporator (Heidolph, Germany) at constant temperature of 40 °C.The dry extract was chemically characterized by the presence of alkaloids, flavonoids, tannins, coumarins, cardiotonic glycosides, saponins, triterpenes/steroids, and quinones.The results were consistent with a previous report (Castro, Franco, Diaz, 2021).
Stability testing
The reactivity of the extract was evaluated against different conditions, according to the recommendations of the ICH and the WHO on stability studies for drugs (ICH, 2003;WHO, 2008).Table I specifies the parameters that were considered and the determinations that were made for each of the dry extract samples (Wu et al., 2009).
The dry extract was conditioned in type I amber vial bottles.All the vials were closed, one part of them was stored under refrigeration, another part of them was placed under room temperature conditions and the rest were placed under temperature stress conditions.(See Table I).Each bottle corresponds to one sampling unit.For each of the sampling intervals and conditions, three samples were removed to analyze the response variables in triplicate and have a fourth, in case a reanalysis was needed (Wu et al., 2009).The response variables established for the study were: organoleptic characteristics, pH, biological evaluation (NO inhibition), and the total phenolic content (TPC) (ICH, 2003;Wu et al., 2009).
Organoleptic characteristics
Various organoleptic characteristics were considered, such as the appearance, color, smell, and texture of the ethanolic extract of A. peruviana seeds.These characteristics were analyzed in each of the sampling times of the stability study (ICH, 2003;Wu et al., 2009).pH A stock solution of A. peruviana seeds extract (1×10 5 μg/mL) in DMSO was prepared.A portion of this solution was diluted with deionized water to obtain a solution of 4000 μg/mL; the pH of this solution was measured using an OHAUS starter 3100 potentiometer.Measurements were performed on all samples at each of the sampling times of the stability study.
Determination of total phenolic content (TPC)
Total phenolic content was determined using the Folin-Ciocalteu's method with slight modifications (Mejia et al., 2020).Briefly, 30 μL of different concentrations of the extracts under study were mixed with 150 μL of a Folin-Ciocalteu solution (0.1 M) and 120 μL of a sodium carbonate solution (7.5%).The mixture was incubated for 2 hours and the DO 760 was determined in a microplate reader (Multiskan GO, Thermo).The results of TPC were presented as milligrams of gallic acid equivalent per gram of dry extract (mg GAE•g-1).
Effect on nitric oxide production and cell viability RAW 264.7 macrophages were maintained routine in DMEM enriched with 10% inactivated FBS at 37 °C and 5% CO 2 .The effect on the release of nitric oxide (NO) was determined as a function of the accumulation of nitrites in the culture medium using the Griess reaction.Briefly, RAW 264.7 cells were plated in 96-well plates (2×10 4 cells/ well), incubated for 48 hours, and treated with the higher non-toxic concentration of A. peruviana extract.Thirty minutes later, cells were stimulated with LPS (1 μg/ml) and incubated for 24 h.Later, the medium was collected and stored at -20 ºC for the subsequent quantification of nitrite.To determine the effect on the viability, MTT solution (0.25 mg/mL) in fresh medium was added to the cells and incubated for 4 hours, after which the medium was carefully aspirated, and the resulting formazan crystals were dissolved in 100 μL of DMSO.Optical density (DO 550 ) was determined using a microplate reader (Multiskan GO, Thermo), and the percentage of cell survival relative to the control group was calculated.NO concentration was determined by measuring the amount of nitrite in the cell culture supernatant using the Griess reaction (Green et al., 1982), equal volumes of Griess reactive and supernatants were mixed and the absorbance of samples was determined in a Multiskan GO microplate reader (DO 550 ), the concentration of nitrites was calculated using a standard NaNO 2 curve (1-200 μM).For all experiments, fresh culture medium was used as blank control, and 1400 W (10 μM), a selective inhibitor of iNOS, was used as positive control.
Compatibility study between the dry extract and excipients
Different binary mixtures of the dry extract of A. peruviana and excipients were evaluated to determine their compatibility.The excipients were selected considering a composition of a microemulsion formulation and their availability.Samples of 300 mg of dry extract were mixed with each one of the excipients at a 1:1 ratio.The binary mixtures between the dry extract of A. peruviana and the required amount of excipient were placed in type I amber glass bottles.Finally, 20 mg of water (5%) were added to each of the samples and were mixed for 1 minute.to facilitate interactions between excipients and the extract.Subsequently, the samples were placed in a circulating air oven at a temperature of 50 °C for 15 days.As a negative control for degradation, the same procedure was carried out for the extract without excipients, placing a sample at 50 °C and another at 4 °C, to determine the true effect that excipients could have on degradation (Bernal, Ramos, Baena, 2019;Kopelman, Augsburger, 2002;Serajuddin et al., 1999;Sims et al., 2003;Wu et al., 2009;Wyttenbach et al., 2005).
Upon completing the 15 days of study, each one of the samples was analyzed to find possible incompatibilities in the appearance and the compatibility was determined by evaluating the thermal behavior by means of DSC, in a Shimadzu DSC-60 equipment.During this process, they were subjected to heating starting with an approximate temperature of 25° C until reaching 200 °C, with a heating ramp of 10 °C/min.The samples were placed in an aluminum sample holder, covered, and hermetically sealed under a nitrogen atmosphere at a flow rate of 50 mL/min.The results were analyzed using the TA Analysis Software program.
Statistical analysis
Assay results of three independent experiments were expressed as the means ± SEM and analyzed using oneway analysis of variance (ANOVA) with post-hoc Dunnet or Tukey comparison.P-values < 0.05 were considered statistically significant.
Organoleptic characteristics
Dry seeds of Ambrosia peruviana were macerated with ethanol until 51.3 g of total extract (yield of 12.73%).The ethanolic extract of A. peruviana seeds showed dark green color as shown in Figure 1 Regarding the organoleptic characteristics observed at each time of the stability test under the different conditions, the extract that must keep under refrigeration (4 °C) or at room temperature (25 °C) had not significant changes; however, at 60 °C had slight changes in the color at the final of the study.In weeks 12 and 24 of the stability study, as shown the Figure 1 B, the A. peruviana extract at 60ºC, presented dark brown coloration.
There are pigments in plants to which their coloration is attributed; among these could be found anthocyanins, carotenoids, and chlorophyll; the latter oversees providing the green coloration to the various plant species.However, the color can be affected by different factors such as light, pH, time, and temperature.Some authors indicated that the conversion of chlorophylls to pheophytins represents the main cause of the loss of green colour in plants.When they are subjected to heat or acidic conditions, said coloring of the vegetables turns into a brownishgreen color (Ahmed, Kaur, Shivhare, 2002).
During the degradation process, hydronium ions can transform chlorophyll into pheophytin by substituting magnesium ions in the porphyrin ring.Figure 2 shows the degradation process of chlorophyll; when heating a plant extract, it may change in color to a brown-green that indicates the degradation of chlorophyll to pheophytin, which is subsequently degraded to pheophorbides which are brown compounds (Ahmed, Kaur, Shivhare, 2002).
pH At the beginning of the stability test, the A. peruviana seed dry extract had a pH of 5.35 and was maintained at an interval of 5.2 to 5.6 as is observed in Figure 3.When these pH values obtained throughout the stability study were reviewed, no statistically significant differences were found between them at the different temperatures used and the different analysis times.
Total phenolic content (TPC)
The Folin-Ciocalteu method was used to quantitatively determine phenolic compounds, based on the reaction of phenolic groups with the Folin-Ciocalteu reagent, a mixture of phosphomolybdate and phosphotungstate that for ms the yellow phosphomolybdotungstate acid which, when it interacts with phenolic groups, reduces it causing a blue coloration.These blue pigments have a maximum absorption according to the quality and/or quantitative composition of phenolic mixtures in addition to the basic pH of the solution, which is generally obtained by adding sodium carbonate (Cicco et al., 2009).To determine the influence of the storage conditions on the A. peruviana seeds extract, the total phenolic content was evaluated at different temperatures and times, protecting the samples from light by using amber flasks to prevent this factor from having a significant influence on the extract.
The content of total phenolic compounds throughout the stability study shows that when evaluating the time at the stipulated storage conditions for the A. peruviana seeds dry extract, the degradation of these compounds is significant over 4 weeks (Figure 4).The extracts stored at 4°C and 25°C have similar behavior (reducing total polyphenols to 18.4 and 20.3%, respectively, of the initial concentration), while at 60 °C the decrease in phenols content in the extract is significantly greater (47.7%).This behavior was previously described in extracts with high phenolic compound content, which presents a significant decrease in phenols after 30 days of storage at temperatures of 6 °C and 23 °C, which are close to those used for us in this assay.This is in addition to agreeing with the authors regarding the fact that the higher the temperature, the greater the Yuri Palacio, Jenny-Paola Castro, Valquiria Linck Bassani, Luis Alberto Franco, Carlos-Alberto Bernal reduction of total phenols in the extract (Srivastava et al., 2007).
The presence of phenolic compounds in plants attribute antioxidant properties due to the ability to capture free radicals due to the presence of the hydroxyl group.However, the stability, biosynthesis, and degradation of this type of compounds can be affected by various factors such as storage conditions and light that can negatively influence the quantity of phenolic compounds present in the plant (Lattanzio, 2003;Valenzuela et al., 2014).
The Folin-Ciocalteu method is widely used to determine the effect of the storage conditions due to the possibility that some of the degradation products could react with the Folin-Ciocalteu reagent, affecting the content of phenolic compounds (Camelo, Sotelo, 2012).In plant species, in addition to bioactive compounds, there are also enzymes such as polyphenol oxidases (PFOs), which are metalloenzymes present in the different organs of plants (roots, seeds, leaves and fruits) and which are responsible for the oxidative degradation of polyphenols, also called enzymatic browning.
The degradation of phenolic compounds by enzymatic oxidation or enzymatic browning may partially explain the change in the color of the extract under heat stress conditions, as phenolic compounds oxidize and polymerize to a brown color (Morante et al., 2014).On the other hand, the possibility of simultaneous degradation of the phenolic and chlorophyll yielding a color change of the extract from dark green to brown can not be ruled out; Regarding the influence of temperature and exposure time on the stability of phenols, significant degradation was observed at high temperatures (60 °C), which increased with exposure time in this condition (12 and 24 weeks).Other authors have already observed this relationship between exposure time and temperature in the degradation of bioactive constituents (Camelo, Sotelo, 2012;Uurrea et al., 2012), corroborating the close relationship between them.
Biological evaluation
The Griess method is a colorimetric test by which organic nitrites are detected.It is carried out using two solutions (A and B).Reagent A consists of sulfanilamide with the addition of a strong acid and reagent B of N (1-naphthyl) ethylenediamine.(Malinski, Mesaros, Tomboulian, 1996).The Griess reaction is based on the formation of azo dyes.The test consists of two reactions, which result in the formation of a colored reagent.In the first reaction the nitrite reacts with sulfanilamide under acid conditions (diazotization reaction) and at the second stage a diazotization product reacts with N-(1-naphthyl) ethylenediamine (diazonium coupling) to form an azo derivative dye.This azo derivative is readily quantifiable by means of a spectrophotometer.Due to its simplicity, speed and satisfactory sensitivity, the Griess reaction has become a widely used colorimetric method for the determination of nitrite in various matrices (Malinski, Mesaros, Tomboulian, 1996).
The biological activity shown by the extract of A. peruviana (15 μg/mL) was conserved during the time of the stability study for the temperatures of 4 ºC and 25 ºC, while the extract subjected to high temperatures (60ºC) showed a significant reduction of the inhibition of NO production in macrophages which is evident from 8 weeks where the inhibition of nitric oxide production has decreased more than 50% (Figure 5).
Macrophages play a critical role in the immune response, its activation with LPS releases large amounts of nitrites which are essential precursors of cytotoxic agents (Hibbs, Vavrin, Taintor, 1987;Nathan, 1992).The excessive production of NO by macrophages during a chronic inflammatory response causes disturbance in essential molecules such as membrane lipids, nucleic acids, proteins, which leads to oxidative stress affecting cellular homeostasis.Hence, the development of products that have substances to prevent the overproduction of NO has become the objective of research to treat chronic inflammation (Desai, Park, 2005;Taira, Nanbu, Ueda, 2009).
Compatibility study between the dry extract and excipients
Figure 6 shows the thermogram of the extract stored at different temperatures (4 and 50 °C).In these curves, two endothermic transitions were detected between approximately 130 and 150 °C.It is generally said that natural products whose composition consists of a great variety of substances, such as the extract under study, present similar thermal behavior (Bernal, Aragón, Baena, 2016, 2019).It should be noted that this behavior is maintained at both temperatures, with the difference that at 50 °C the endothermic transitions are smaller in size compared to what is observed at 4 °C, concluding that the storage temperature influences the thermal behavior of the extract but not significantly.However, the aim is to incorporate the extract into a microemulsion, therefore, a series of excipients were selected according to the chosen pharmaceutical form, hoping that when mixing with the extract the thermal behaviors of both would be maintained, indicating that they are compatible and that therefore both it is possible to use them in a formulation (Uurrea et al., 2012).
The thermal behavior of the extract, the respective excipient and the extract-excipient combination are compared in the DSC curves.For them to be considered to be compatible, the thermal behavior of the extract and the excipient in question must be maintained.The example in Figure 7 shows the thermal behavior of the extract, the excipient 1 (Castor Oil) which is highly thermostable as it does not present peaks, and the combination of both, from which it can be said that castor oil is compatible with the extract, because the mixture maintains the two endothermic transitions characteristic of the extract, and that in this case the excipient that does not present thermal transitions in this temperature range.It does not alter the behavior of the extract when mixed with it.After carrying out the analysis of the different thermograms of the mixtures and comparing them with the thermal behavior that the extract and the excipient presented separately, it was found that within the working conditions used during the study by DSC there is no chemical incompatibility between the extract of A. peruviana seeds and a large part of the selected excipients, (Data not shown).Here the thermal behavior of the components is maintained separately or they suffer variations such as widening or displacement of endothermal peaks, which are not significant since this type of phenomenon can be typical of the decrease in the purity of each component due to the mixture (Kiss et al., 2006;Ranjan et al., 2012).
As an example of interactions, Figure 8 shows the thermogram of the extract and excipient 4 (sodium lauryl ether sulfate) which presents an endothermic peak between approximately 100 and 120 ° C and that when mixed with the extract, the disappearance of this thermal peak is observed, which may indicate a possible interaction among them.Nevertheless, when observing the thermograms of excipients 2, 4, 5, 12, 14, 16, and 20 (Data not shown), it is possible to appreciate marked differences in the mixture curve, which probably lead to interactions among them.These differences may be due to various factors, among which we find the disappearance or appearance of peaks in the curves, due to possible chemical extract-excipient interactions when the mixture is heated.It should also be noted that in some of the mixtures previously mentioned as incompatible, there are significant widening and displacement of the peaks of both the extract and the excipient.Said thermal behavior may indicate an alteration in the characteristic of some of the components, and that could compromise the stability and pharmacological activity of the extract.However, most excipients turn out to be compatible for this reason; it can be said that they are considered potential candidates to be included in microemulsion formulations that contain the extract of A. peruviana seeds.The results of the excipient compatibility are shown in the Table II.
CONCLUSION
The preformulation studies of the ethanolic extract from seeds of Ambrosia peruviana all suggested it was stable at temperatures of 4 °C and 25 °C during the 24 weeks, presenting losses of approximately 21% of the initial phenolic content.On the other hand, it was unstable at 60 °C losing approximately 50% of the initial phenolic content after 12 weeks.Moreover, the antiinflammatory effect evaluated on nitric oxide production and cell viability demonstrated the excellent activity of the extract, which did not change at 4 °C and 25 °C.The compatibility study among the dry extract and excipients showed that the extract was compatible against different excipients: oils, surfactants, cosurfactants, antioxidants, preserving agents, stabilizing agents, and vehicles.This study is the first step for formulating Ambrosia peruviana seeds ethanolic extract as a standardized microemulsion dosage form.
FIGURE 1 -
FIGURE 1 -Ambrosia peruviana seeds dry extract.A. at time zero.B. At time 12 and 24 weeks in different conditions.
A, characteristic odour and appearance of viscous fluid, what makes handling difficult.
FIGURE 3 -
FIGURE 3 -Changes of pH of the extract in the stability study.
FIGURE 4 -
FIGURE 4 -Changes of the total phenolic content in Ambrosia peruviana seeds extract during storage at different temperatures.**P<0.01;****P<0.0001statistically significant one-way ANOVA and Dunnett's test against the 4 °C group.
FIGURE 5 -
FIGURE 5 -Inhibition of NO production by Ambrosia peruviana seed extract during storage at different temperatures for 24 weeks.
FIGURE 6 -
FIGURE 6 -DSC curves of Ambrosia peruviana seeds extract stored at different temperatures.
TABLE I -
Conditions and variables of the stability | 5,600 | 2023-05-22T00:00:00.000 | [
"Medicine",
"Chemistry"
] |
Nonvolatile programmable silicon photonics using an ultralow-loss Sb2Se3 phase change material
An ultralow-loss optical phase change material enables reversible programming of the flow of light in silicon photonics.
INTRODUCTION
The birth of coherent nanophotonic processors, photonic tensor cores, quantum computing, and neuromorphic networks signifies a large paradigm shift toward emerging optical information platforms (1)(2)(3)(4)(5)(6). Postfabrication programming of devices is one of the most desirable functionalities for agile reconfigurable photonic functionalities (7,8). Despite the great success of implementations based on cascaded in terferometer meshes (2,(9)(10)(11)(12), there are inherent strong limitations in footprint scalability and volatility. Therefore, the development of new concepts and technologies is of extreme interest.
The fundamental benefits of using nonvolatile phase change ma terials (PCMs) in reconfigurable photonics have resulted in their extensive exploration for photonic modulation and resonance tuning, with Ge 2 Sb 2 Te 5 (GST) (13)(14)(15) and, more recently, Ge 2 Sb 2 Se 4 Te 1 (GSST) (16) being the most considered materials. The multiple non volatile states these materials offer provide unparalleled energy per bit operation in a highly stable platform. Commercial GST phase change memory has been shown to be stable over 10 12 write cycles (17). Both materials operate based on a large change in complex refractive index ñ = n + ik between their crystalline and amorphous phases. Despite improvements in materials design, the absorption losses in one or both phases of current PCMs prevent optical phase control independent of changes in the amplitude of propagated light in the telecommunication band, severely limiting phase modulation schemes.
In several recent studies, the antimonybased chalcogenides Sb 2 S 3 (18,19,20) and Sb 2 Se 3 (21) have been identified as a family of highly promising ultralowloss PCMs for photonic applications. The materials exhibit no intrinsic absorption losses (k < 10 −5 ) in either phase over the telecommunications transmission band and show low switching temperatures around 200°C for Sb 2 Se 3 and 270°C for Sb 2 S 3 (21) while remaining nonvolatile at operating temperatures. The transition between crystalline and amorphous phases changes the arrangement of the chemical bonds in the material, which in effect results in the change of optical properties such as the complex index of refraction. Furthermore, the proximity of their refractive index to that of silicon allows straightforward direct integration of PCM patches onto standard silicononinsulator (SOI) integrated photonic platforms with excellent mode matching to the SOI waveguide.
Here, we demonstrate the exceptional capabilities of the PCM Sb 2 Se 3 for use in ultralowloss optical phase control of photonic in tegrated circuits (PICs). To achieve this, we make use of 23nmthin patches of Sb 2 Se 3 deposited on top of a 220nm SOI rib waveguide, where the thickness of materials is chosen to maintain a single mode of propagation. An optical phase shift of the propagating wave is induced by changing the material of the PCM from a crystalline state to an amorphous state. In our studies, following the example of optical storage media, we start from a fully crystallized PCM as this provides a uniform background, fast writing speeds, and stable amorphous written areas. Switching of individual pixels in the PCM is achieved using tightly focused optical pulses from a diode laser operating at 638nm wavelength (see Materials and Methods). Sb 2 Se 3 is an emerging new material in the domain of PCMs with fundamental advantages for integrated photonics over established PCMs. To validate our approach and quantify the optical phase shifts induced by switching this PCM, we start our investigation by char acterizing its response in a standard MachZehnder interferometer (MZI) phase shifter configuration. The integration of Sb 2 Se 3 PCM patches in a standard MZI is of key importance for conventional reconfigurable platforms (5-10) as it will allow replacing volatile active phase shifters with nonvolatile optical phase control.
Following the precise characterization of the optical phase shift associated with the PCM pixels in the MZI, we demonstrate the proof of principle of a programmable multiport router based on writing patterns of weakly scattering perturbations onto a multimode inter ference (MMI) device, such as the one shown in Fig. 1A. Figure 1B shows a schematic layout of this experiment and highlights our gen eral approach of optical writing of pixels using a microscope. Fol lowing on earlier proofofprinciple studies (22,23), the capability of writing nonvolatile pixel patterns in a multiport waveguide signi fies a breakthrough in freeform control over the flow of light inside integrated photonic circuits. Currently dense meshes of singlemode devices are being considered as the main strategy for programmable optical circuits; however, this is expensive in both the number of components and device footprint. The results presented here are a first step toward a new family of programmable multiport elements, which could open up entirely new architectures for photonic inte grated circuits (23).
Switching of Sb 2 Se 3 patch on a Mach-Zehnder interferometer
In our work, we use a standard 220nmthick SOI platform where 120nmthick rib waveguides are produced by partial etching of the silicon layer. The refractive index values of Sb 2 Se 3 in its two states (n amorph = 3.285 and n cryst = 4.050, respectively, for amorphous and crystalline states) are close to that of silicon (n Si = 3.48), resulting in very similar mode profiles and hence low insertion losses between the different types of waveguide configurations (see table S1). Fig ure 1 (C to E) illustrates the geometry of the silicon photonic plat form under study and includes the mode profiles for the different configurations. From mode calculations, we found that an amor phous single pixel in an otherwise crystalline Sb 2 Se 3 layer changes the effective refractive index of the waveguide, n eff , by an amount n eff = −0.072 (21).
By selective switching of the PCM in one arm of an asymmetric MZI, the optical phase shift is translated into a change of the trans mission function of the device. The first demonstration is shown in Fig. 2A, in which we achieve optical phase tuning of an MZI with a 125mlong patch of Sb 2 Se 3 . Progressive tuning of the optical phase is achieved by amorphizing a series of spots in a continuous line along the PCM patch. Spectra were taken every 25 pixels, where each pixel is spaced by 1000 nm along the MZI. The figure shows a suc cession of vertically offset recorded spectra (gray) from the initial state (red) to the fully switched state (black), followed by a reset to the final state (blue). A detailed view shown at the bottom of Fig. 2A shows the initial, set, and reset states, with the insertion loss for each spectrum normalized to the initial spectrum. The insertion loss re mains unchanged during this process, highlighting the lack of any losses induced by the amorphization or recrystallization. Changes in the effective refractive index n eff along the patch are expected to result in very small reflective losses of <0.02 dB per interface (see table S1). A reversible 2 phase shift was obtained as a result of 100 separate crystallization/amorphization pulses. Regardless of the small imperfections introduced by the motorized stage (backlash) during recrystallization, the full 2 range was set and reset with a 0.1nm offset, which results in an accumulated error within 2.5% of the full range. This measurement demonstrated that a 750nm pixel size results in a resolution of 0.02 shift, allowing very fine phase control.
As the induced optical phase change scales linearly with amor phized distance and given the ultralow loss of the material, very large phase tuning ranges can be readily achieved by using longer PCM structures. Figure 2B shows a 10 phase shift obtained by switching a 250m length of PCM patch, with no measurable change to in sertion loss or modulation depth. The effect of the PCM on inser tion loss and extinction of the MZIs is explored in more detail in fig. S2, which shows that for PCM patch lengths of up to 250 m, no additional rebalancing of the MZI is required to compensate for losses in the PCM. Apart from using the focused laser spot size for calibration of the switched pixel, the induced optical phase can be finetuned by small changes in the laser power. The phase shift was tuned to 0.04 per pixel in the experiment of Fig. 2B by a small in crease in optical peak power of the diode write laser. This average phase shift per pixel was calculated as the total phase shift divided by the number of pixels. Total phase shift was obtained from the spectral shift of the fringes in the MZI response.
To complement the broadband measurements, we performed additional narrowband measurements on an MZI with a shorter, 50m patch length as shown in Fig. 2C. Starting at the wavelength of 1553 nm, we tuned the MZI from the maximum transmission (set to 0 dB) to the minimum (−17 dB) and subsequently reset part of this device by recrystallization of a 20m length of the PCM patch (red curve). The curve shows the individual levels in transmission induced by switching subsequent pixels in steps of 500 nm along the PCM. An asymmetry between the first write and reset scans is ob served, which is attributed to a burnin effect of the initial film. A second amorphization scan (second blue area) follows more closely the reset scan, where we notice a small but noticeable drop in the maximum extinction of the device from −18 to −15 dB, indicative of small changes in the material upon repeated switching of the PCM layer, changing the sensitive balance between the MZI arms (see also fig. S2). Overall, this result highlights the ability to define stable multilevel switching processes for highly dense and accurate optical phase control with very fine quantization.
Next, the endurance of the phase change was probed by repeatedly switching a single pixel of the MZI. We used a pixel at the −13.5dB point of the MZI response curve on the steepest part of the slope, where the sensitivity to individual pixels is high. By repeatedly cycling between the amorphous and crystalline phases in a single spot, the transmission modulation gives a good indication of the lifetime of the phase change. Figure 2D shows the transmission in each phase for the first 600 switches, with red dots corresponding to the crystalline and blue dots to the amorphous material phase. A longterm drift of the device transmission, due to environmental fluctuations and material changes, was observed, and we used a 50point moving average to normalize the drift, as shown in the bottom panel of Fig. 2D. The first 350 full setreset cycles resulted in a constant transmission change, with a small increase seen toward the 100th cycle. This increase was observed in most tested devices and is attributed to the settling of the switching dynamics after the first few burnin cycles of the PCM, resulting in a slightly increased responsivity. A reduction of the switching contrast below 50% of the initial value is seen after around 500 cycles. Another example of an endurance experiment is shown in fig. S4. We note that the en durance seen in our current work is a factor 10 lower than that ob served for Sb 2 Se 3 on planar films (21), which is attributed to the somewhat more challenging thermal environment of the PCM on the waveguide and may require further improvements in ther mal design.
Digital programming of MMI using precalculated pixel patterns
Having successfully demonstrated that an optical phase shift can be induced by switching thin Sb 2 Se 3 patch on top of an MZI, we moved on to demonstrate a new approach for a programmable optical router based on an MMI patch. MMIs are of interest for their small foot print as well as for their substantially reduced sensitivity to the envi ronment when compared to MZIs. The device principle is based on the concept of wavefront shaping, where a distribution of weak per turbation pixels is used to effectively steer the light with very low loss (22,23). The perturbation induced by switching of the PCM lay er is sufficient to apply this device concept in a practical scalable device configuration. Furthermore, a high level of control is achieved, allowing us to independently set and reset pixels spaced at 800nm pitch in complex patterns.
A 1 × 2 MMI device of 33 m × 6 m, covered with a 23nm crystalline Sb 2 Se 3 film, was simulated using a twodimensional (2D) finitedifference timedomain approach, and a pixel perturbation pattern was optimized using an iterative scheme (22). In this scheme, we switched every pixel of the MMI individually and investigated the resulting MMI outputs. Switched pixels were kept if they im proved the target response and were reset when they did not result in an improved target response. Through this iterative process, the optimum combination of pixels in the region between the input and output ports was determined, and we therefore maximized the transmission from one of the outputs. Repeating this optimization two times resulted in a pattern with high singleport transmission. Figure 3 (A and B) shows the calculated intensity in the device be fore and after patterning. The designed pattern was subsequently written onto an experi mental device. Figure 3 (C and D) shows the fabricated MMI with the same dimensions and cladding as the simulated ones before phase change (C) and after the perturbation pattern was written (D). A residual structure can be seen in the MMI in the unperturbed state (Fig. 3C), which is caused by the crystalline domains of Sb 2 Se 3 in the crystalline (unswitched) state. In Fig. 3D, a very good agreement is seen in the pattern registration, with a small ~0.5° tilt present in the experimental pattern due to limitations in the alignment of the set up. Each pixel was written sequentially in columns and rows from the input to the output. Figure 3 (E and F) shows the simulated (dashed curves) and ex perimental (solid lines) transmission in the top and bottom outputs for both the set (amorphization) and reset (crystallization) steps, where the singleport transmissions were obtained by normalizing to the sum of the two output ports. The absolute MMI singleport transmissions before writing are −30.6 dB (top) and −30.8 dB (bot tom) compared to −25.9 dB for a straight waveguide. After writing the full pixel pattern, a 92%:8% splitting ratio was found between the top and bottom outputs, respectively. Running the final pattern from the 2D optimization sweep in a full 3D simulation for the fully patterned MMI provides a 97%:3% splitting ratio, as shown by the dots in Fig. 3E. The 3D simulation results indicate that optimization using a 2D effectiveindex model provides a good approximation despite the complex 3D geometry of the perturbation positioned on top of the waveguide. The 5% difference between experimental and simulated results is attributed to experimental limitations in pattern registration onto the device, which are as yet not fully understood. The device reset was done pixelwise in the same way as the write sequence, i.e., from start to end of the pattern; therefore, the curves in Fig. 3 (E and F) are not the inverse of each other. We point out that the reset was not perfect in all areas. The thermal properties at the edges of the MMI are different due to the surrounding insulating ZnS:SiO 2 capping layer, which results in different switching dynamics.
Eventually, the material breaks down at the edges of the MMI, which results in nonreversible switching. This can be mitigated by implementing smoother interfaces and also by using a variable laser power depending on the local topology.
Iterative optimization of MMI transmission
Next to patterning the MMI device using a calculated pixel pattern, we tested an alternative scheme of iterative optimization of the ac tual device. In the iterative approach, we switched individual pixels and tested their effectiveness on the target transmission. For every pixel, a test was performed to see whether switching increases the transmission at the output port. If a pixel contributed positively to the target, its switched state was preserved; otherwise, it was erased. Whereas a singlepass optimization already resulted in a notable power splitting between the two ports, additional passes are shown to further improve the target transmission function. Figure 4 (A and B) shows results for bottom (blue) and top (red) output ports where only the accepted pixels are shown, and every data point corresponds to a switched or reset pixel in the MMI. Figure 4 (C and D) shows the same data on a logarithmic scale and normal ized to the original port intensities of the unperturbed device, T 0 . Raw port output signals are subject to intrinsic variations in grating efficiency and collection fibers, therefore normalizing the signal looks at relative changes compared to the unperturbed device. The entire device was optimized in one pass (pass 1), followed by a second pass, which started again at the beginning of the device and which reoptimized the first 20 m of the MMI (pass 2). Figure 4C shows an increase of singleport transmission of the bottom output by 2 dB, accompanied by a 6dB reduction of the top output. The total transmission, defined as the sum of the unnormalized T top + T bottom , shows an overall 0.5dB increase in insertion loss caused by writing the pattern. The changes induced by writing the pixel pat tern are reversible, and the original device transmission is retrieved by resetting all pixels in the device as shown in Fig. 4 showing the complete iterative writing process and the device reset are presented in the Supplementary Materials. Both precalculated and iterative optimization patterns result in a target performance exceeding 8dB extinction between the transmission from the two ports. The impact of the patterning on the total device throughput is not strongly affected, with only up to 0.5 dB additional loss for all devices under study. In some cases, an improvement in device trans mission was seen, as shown for example in fig. S6. Such a small im provement in overall device throughput can be attributed to the effect of iterative wavefront shaping, which compensates for some of the imperfections of the original device, as seen in earlier studies (22).
Reversibility of write and reset of MMI pixel patterns
To show that the pixel pattern writing approach is both reconfigurable and freeform, the same pattern was written into an MMI multiple times, with a full reset between each pattern for both the simulated pattern and the mirror image of the pattern in the vertical direction, which, due to symmetry, guides light toward the bottom output. Figure 5A shows the MMI before (1) and after (2) the first write cycle. To improve the visibility of any changes with respect to the original MMI patterns, images of all set and reset states were pro cessed by differencing with the original MMI, as shown in Fig. 5B. A good repeatability was shown between each pattern, with the reset state showing no memory of the patterns written. After several switching cycles, incremental damage to the Sb 2 Se 3 film occurred at the edges of the MMI. The damage appears from the corners of the device, where the lower local thermal conductivity of the surrounding ZnS:SiO 2 capping layer and the abrupt interfaces result in a higher maximum temperature, making the Sb 2 Se 3 film more prone to de lamination or void formation. These effects can be compensated by a variety of methods such as reducing the pulse power for these ar eas or by reducing the actively written area of the MMI. Figure 5B shows the transmission of both outputs for all the perturbations when writing and resetting the patterns shown in Fig. 5A. A repeatable transmission change was seen for both orientations, which was not strongly affected by the increasing presence of the damaged areas (see Fig. 5A). The splitting ratio was higher for the inverse pattern, suggesting either a small misalignment between the pixel pattern and the MMI or a nonsymmetrical design due to fabrication tolerances in the Sb 2 Se 3 film.
DISCUSSION
The results presented here unequivocally show the capabilities of the new phase change technology in providing lowloss programmable optical phase control in PICs. This approach is expected to open many new applications in postfabrication device tuning, program mable weight banks, unitary matrix operations, and, ultimately, all optical fieldprogrammable arrays (10,23). Nonvolatile PCMbased approaches may hold an energy advantage compared to active de vices, which require driving voltages to maintain a configuration and hence could offer opportunities for reduced power consump tion PICs. The use of nonvolatile programmable MZI with large optical phase shifts exceeding 2 is of interest. Referred to as optical true time delay (OTTD), programmable optical path lengths are of interest in microwave photonics (24), optical fast Fourier trans forms (25), Fourier transform sensors (26), and integrated quantum circuits (27). In microwave photonics and emerging terahertz (6G) applications (28,29), OTTDs are used for beam formation using phase array antennas and in optical communications for signal synchro nization, equalization, buffering, and time division multiplexing. In quantum optics, precise tuning of optical path length differences is critical for maintaining high multiphoton coherence (30). Our current results are aimed at applications in programmable and reconfigurable photonic circuits, which do not rely critically on switching speed or extreme endurance. Applications requiring ex tensive cycling and/or highspeed switching, such as memory cells, displays, or other adaptable elements, will require further extension of the limits of performance of these materials integrated into highly optimized device configurations. For field use, electrical switching through Joule heating may be possible. To mitigate the complexity of switching each individual pixel electrically, preselected deposited pixel arrangements could be switched with fewer connections sacri ficing resolution for lower interconnect count. We envision that im portant aspects enabling the mass commercial use of this technology such as the reproducibility, switching endurance, and optical and electrical switching will be addressed in followon studies, as has been the case with decades of work in optimizing materials like GST.
In conclusion, we have demonstrated a new low loss reconfigu rable approach for optical routing in an integrated silicon photonic device based on pixel patterns written on an MMI. The new PCM Sb 2 Se 3 allows the decoupling of optical phase control from ampli tude modulation seen in the conventional PCMs. The advances in device footprint and energy consumption of this approach compared to conventional cascaded switch fabrics could enable a range of complex photonic circuits needed for applications such as onchip light detection and ranging, photonic quantum technology, artificial intelligence hardware, or optical tensor cores of the future while pro viding a powerful postfabrication programming technique for high volume PIC ecosystems. Our demonstrated technique provides a general approach that could be easily extended to larger devices and could ultimately achieve a platform for a universal optical chip technology.
Materials
The devices shown in this work were fabricated using an industrial deep ultraviolet (UV) scanner on a 220nm SOI platform with a 120nm etch depth. The UV scanner was used to open windows above the MMI and MZI regions, and 23nmthin Sb 2 Se 3 films were deposited using lowtemperature radio frequency (RF) sputtering. Following removal of the photoresist, the whole chip was then capped with a 200nmthick layer of ZnS:SiO 2 , using RF sputtering.
Methods
A 150mW, 638nm wavelength singlemode diode laser was used to crystallize and amorphize Sb 2 Se 3 , using pulses of 50 ms at 19 mW and 400 ns at 35 mW, respectively. A 0.42 numerical aperture (NA) 50× objective was mounted to a 3D piezo stage, normal to the devices under test, which was for focusing the laser and taking optical images with a white light source and chargecoupled device camera. Using the piezo stage, the objective was moved to create the pixel patterns with high reproducibility. More details on the optical setup are pre sented in fig. S1. | 5,619.6 | 2021-06-01T00:00:00.000 | [
"Physics",
"Engineering",
"Materials Science"
] |
COVID-19 and food security in Africa: Building more resilient food systems
The COVID-19 pandemic has exposed the fragility of our food systems. Despite increased efficiencies in producing and supplying large volumes of food, our current food systems have generated multiple adverse outcomes comprising high greenhouse gas emissions, persistent hunger, and livelihood stress for farmers around the world. Nowhere else than in Africa have large numbers of people experienced more acutely these adverse shocks emanating from our food systems. Thus, building more resilient African food systems, which take a radical change of direction, is fundamentally a matter of survival. While there is broad consensus around a need for transformational change in food systems, what that entails is not always clear, and there are divergent views amongst experts on how to re-orient research priorities and agricultural solutions in ways that effectively address hunger and inequality while also protecting agrobiodiversity and the environment more broadly. This article engages with this debate and proposes an agricultural research for development agenda in Africa that balances technology transfer with realigning societal values, institutional arrangements, and policy decision-making towards the realization of greater sustainability and inclusive outcomes.
Introduction
Various actors point to threats of a looming global food crisis 1 due to the impacts of the novel coronavirus . In major African cities such as Nairobi, Kinshasa and Lagos where up to two-thirds of the population rely on the informal sector for their livelihoods, millions of people have been left without income to purchase food due to the abrupt loss of jobs that often provide daily earnings. In rural areas where agriculture is the main source of people's livelihoods, disruptions to transportation and logistics have made it difficult for producers (farmers, livestock keepers, fisherfolk) to sell their produce and to gain access to agro-veterinary inputs and services.
In Kenya, for instance, a country-wide curfew (7 PM-5 AM) and movement ban in and out of four counties, including Nairobi, have seen smallholder farmers who produce over 70% of the food consumed in the country face high transportation fees to deliver their produce to cities 2 while others scramble to find alternative markets. The effects of these restrictions might result in higher food prices, akin to experiences from the Ebola crisis in West Africa in 2014, which disrupted agricultural supply chains. Today, the COVID-19 pandemic puts a further strain on Africa's agricultural sector which is already facing unfavorable climate change patterns involving a higher frequency and intensity of extreme weather events such as droughts and floods, market and price volatility, and the recent desert locust outbreak in the Horn of Africa.
To address the immediate food security shocks brought about by the COVID-19 pandemic, multiple African governments have introduced relief measures to cushion the poorest and most vulnerable segments of their populations as well as to ensure that producers have affordable access to farm inputs.
Food system resilience paradigms
Beyond tackling the immediate concerns surrounding health and food emergencies, global food security leaders reiterate that the COVID-19 crisis offers an opportunity for decisive collective action towards building resilient food systems that enhance ecological sustainability and equitable outcomes 3 . The COVID-19 outbreak has brought to the fore some of the existing challenges facing our food systems. For example, while current food systems have become efficient at producing and supplying large volumes of food, they have generated multiple adverse outcomes comprising high greenhouse gas emissions, persistent hunger, and livelihood stress for farmers around the world 4 .
Nowhere else than in Africa have large numbers of people experienced more acutely these adverse shocks emanating from our food systems. Thus, building more resilient African food systems, which take a radical change of direction, is fundamentally a matter of survival. African Ministers of Agriculture, speaking on the impact of COVID-19 on food security and nutrition in Africa, have emphasized that developing sustainable and resilient food systems in Africa can address various negative influences beyond providing adequate food, including on public health, youth employment, education, economic development and social well-being 5 .
While there is broad consensus around a need for transformational change in food systems, what that entails is not always clear, and there are divergent views amongst experts on how to re-orient research priorities and agricultural solutions 6 in ways that effectively address hunger and inequality while also protecting agrobiodiversity and the environment more broadly.
For some actors, resilient food systems are productive and efficient, and operate under the principles of climate-smart agriculture and sustainable intensification. Ideal food systems are also envisioned to support the inclusive participation and economic empowerment of especially marginalized food producers and agricultural workers such as through favorable integration into food value chains 7 . For other actors, resilient food systems promote diversified agroecological farming and landscapes, based on the principles of food sovereignty, and are intended to produce diversified outputs ranging from increased dietary diversity from locally produced food, minimal shocks from crop and market failures, and autonomy over how resources are used 8 . Further, it brings greater scrutiny and critical debate to normative assumptions surrounding food security as well as to the political capital and priorities that lock-in place current unsatisfactory practices in food systems 9 .
Both of these food systems resilience paradigms offer a wide diversity of proven agricultural technologies, practices and valuable insights that can enhance the resilience of Africa's producers and food systems. However, agricultural technology uptake amongst producers has often proved challenging and many commendable innovations fail to expand their reach to large numbers of beneficiaries in a manner that is selfsustaining or long-lasting. There are multiple reasons for their limited impact at scale. Among them is that most efforts tend to narrowly focus on generating more produce or income from existing farms but pay relatively scarce attention to the different dimensions of agricultural systems constrains (mainly economic and institutional in nature), or whether innovations necessarily align with the specific constraints, interests and motivations of smallholder producers and other stakeholders 10 .
Building more food system resilience in Africa
Building more resilient food systems in Africa will require reconfigurations that balance technology transfer with realigning societal values, institutional arrangements, and policy decisionmaking towards the realization of greater sustainability and inclusive outcomes 11 . This process will need to pay attention to and support the following elements: • Offer low-cost or cost-effective agricultural innovations and practices that can enhance the resilience of Africa's resource-constrained producers to hedge safely against risks in environments where they are routinely subject to multiple unpredictable shocks and outcomes (e.g., crop loss, market failure).
• Build the agency of individuals and communities to foster ownership in the management and control of such agricultural innovations, as well as to advocate for their own priorities and interests more effectively, beyond the short-term duration of typical agricultural development interventions.
• Engage with decision-makers to advocate for the implementation of strong institutional or policy mechanisms that support context-appropriate agricultural solutions and can enhance resilience in Africa's food systems.
• Support research and learning that informs African food producers and consumers about the value of strengthening food systems resilience in a manner that provides nutritious and healthy food while delivering livelihood benefits to farmers and promoting sustainable agricultural practices.
When taken together, these elements can help to reposition agricultural interventions to enhance food systems resilience impact in Africa. Various actors and initiatives are already working towards this agenda. However, their efforts often face enormous structural constrains and struggle to gain the necessary political and financial support needed to meaningfully expand their impact at scale.
For example, Bioversity International and the CGIAR Research
Program on Climate Change, Agriculture and Food Security (CCAFS) have supported the establishment and capacityenhancement of community seed banks in Uganda, Kenya and Uganda 12 . This work aims to lend greater to farmers' social seed networks, commonly referred to as informal seed systems, which in most parts of Africa supply up to 80% of the seeds grown by smallholder farmers. These community seed banks serve multiple key functions. Among them is in-situ conservation of local plant genetic resources, and access to greater varietal diversity of seeds and planting materials acquired from national and regional gene banks and farmer-to-farmer exchanges. Despite this vital role that informal seed systems play in the production and distribution of a vast majority of the seeds used by Africa's smallholder farmers, they are often overlooked in dominant seed system development endeavors which largely favor the expansion and commercialization of formal seed systems 13 .
Similarly, the Alliance for Food Sovereignty in Africa (AFSA), one of Africa's biggest civil society movements, is involved in an advocacy campaign for agroecology across Africa. Agroecology seeks to (re)design farming systems in ways that maximize agrobiodiversity using a wide range of crops, seed varieties, and farm animals, as part of a strategy to stabilize food supply against climatic variability and seasonal shortages, while building healthy agro-ecosystems 14 . For AFSA, agroecology should focus on building upon Africa's diverse food systems and traditional farming practices, while ensuring that farmers are in control of all aspects of food production. While AFSA has seen increased recognition in some global policy arenas, the movement faces difficulty accessing critical national and regional policy spaces that can facilitate the implementation of some of its agroecology efforts 15 . To them, African agricultural policymaking is biased towards intensifying the production of staple crops, using a narrow range of agrochemicals and improved seeds as evident in multiple Farmer Input Subsidy Programmes 16 .
As Africa's policy-makers grapple with how to meet the food security demands of their nations considering disruptions caused by the COVID-19 pandemic, now is also a time to consider system-wide reconfigurations that can build greater resilience in local and national food systems. Evidence from Africa-based organizations and movements demonstrate that investing in approaches that build the agency of producers and their communities to improve their agricultural practices can guarantee the stable supply of healthy and nutritious food. These efforts can help feed Africa adequately and sustainably, but they will need much greater political and financial support.
Data availability
No data are associated with this article.
Acknowledgements I would like to thank Joab Osumba and Catherine Mungai for helpful comments on an earlier version of this paper.
The author argues for a radical shift in policy and research direction for addressing these pressing challenges and suggests building more resilient food systems that overcome underlying bottlenecks such as: improving public health services, access to education and youth unemployment. The author describes varying approaches to building resilient food systems, through food sovereignty, agro-ecology and through improved productivity and efficiency but argues that these approaches often result in narrowed outcomes that are focused on economic gains.
The author makes a number of recommendations for building more resilient food systems that could withstand future shocks and uncertainties, and that could mitigate the long-term impact of the current global pandemic. The article makes a strong case for shifting attention from technologically-focused innovations to cost-effective, demand driven strategies that consider individuals' and communities' agency, their control and ownership over innovations and an overall more holistic approach to multi-level decision making, learning and education.
The author suggests that a radical or transformational change is necessary for future resilient food systems across the SSA region. The analysis could pay further attention to how substantive structural changes are needed. For example, analysis of financial and political actors that support current and future scalable and technologically-driven initiatives could identify specific investment needs around integrating resilience strategies into food systems.
Considerations for the politics of knowledge and power relations behind agenda-setting and policy formations are also necessary to ensure that strategies are designed to meet the long-term, sustainable needs of both rural and urban dwellers in SSA. An emphasis on the processes involved in building resilient food systems could highlight approaches that include those affected involved in building resilient food systems could highlight approaches that include those affected the most, (for example, those based in the informal labour sector) in collaborative, co-design planning, research and implementation.
Shocks brought on by the pandemic also pose specific risks to different actors within food value chains. Informal food vendors, material suppliers, factory workers (in food processing) for example, are all vulnerable to contracting the virus but who also risk losing their livelihoods without continuing their work in some capacity. Social protection measures are a necessary factor in establishing future food systems that benefit everyone. Further to this, context-specific, intersectional considerations (including gender, sexuality, race, class, ethnicity and able-bodyness, etc.) in the analysis could further the discussion around social and economic inequalities within food systems, and the exacerbating impact of the pandemic on the most marginalized populations. I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard. | 2,989 | 2020-06-26T00:00:00.000 | [
"Agricultural and Food Sciences",
"Economics",
"Environmental Science"
] |
Low-cost Arduino-based Temperature Measuring System
The commercial equipment that carries out the measurement of temperature has a high cost. Therefore, this article describes the development of a temperature measurement equipment, which uses a microcontrolled platform, responsible for managing the data of the collected temperature signals and making available the acquired information, so that they can be verified in real time at the measurement site, or remotely. The construction of the temperature measurement equipment was performed using open platform hardware / software, where performance tests were carried out with the objective of developing a temperature measurement equipment that has measurement quality and low cost.
Introduction
Automatic test and measurement (or data acquisition) [1] [2] systems are used to experimentally evaluate the parameter values of a process, product or experiment.They are different from monitoring systems integrated in process / plant supervision / control because the measured values are not used directly (ie as part of the same system) to automatically adjust the process / plant under test [3].At an industrial level, increased automation in the production process increasingly amplifies the needs for accurate measurement systems for quality control and, consequently, the need for software development.At the scientific level, similarly new experiments and machines (such as particle accelerators, astro-telescopes, or space missions) require an impressive performance of automatic measurements and systems, often beyond the limits of the latest generation [4].
This approach promises access to low-cost sensor-based instrumentation by resource-poor researchers, underdeveloped and developing world laboratories.Historically data acquisition systems have a high cost, both for small and medium companies, as for the academy [5].This article provides a methodology for applying an open source approach to designing and developing a low cost temperature acquisition system that allows data from the monitored process to be accessed remotely.
As an additional resource, because the acquisition system has a remote communication interface, it meets the most current in the industry, called industry 4.0, or interconnected industry, whereby equipment and operator interact using communication resources via the network of data, also called IoT (Internet of Things) [6] [7], which Ray [8]defines as a platform where every day the equipment and processes become smarter, and every day the communication becomes informative.The IoT architecture can be treated as a system that can be physical, virtual, or a joint of the two, which consists of the use of numerous devices such as sensors, actuators, cloud services, communication protocols, developers, etc.This feature stands out in a measuring equipment because providing real-time temperature information assists the operator of the machine whose equipment is installed in decision-making.
Recent work shows that the costs of scientific equipment can be dramatically reduced by applying open source principles to your project using a combination of the open source Arduino electronics prototyping platform as described in section 2.
Fatehnia et al., described an automatic double-ring infiltrater system using the Arduino platform, using hall and water level sensors to perform the measurements.The measurements are validated in relation to existing systems, and the collected data is stored in SD card.
Mesas-Carrascosa et al., used the Arduino platform to ensure best agricultural practices, obtaining real-time information on temperature, air and soil moisture.The system uses the information collected and previously processed in relation to performance models to monitor the crop, and can be accessed remotely through a mobile application.
Ali et al., described a project using the Arduino platform to measure and record data indoors.The system uses several sensors to measure temperature, air humidity, light intensity, CO2 concentrations, voltage, differential pressure and human occupancy of the environment.From the data collected, a research is carried out to investigate the parameters that influence the quality of life in these environments.
Kviesis e Zacepins carried out the monitoring of bee colonies through the acquisition of temperature, humidity and beehive weight signals.All information is obtained and processed using the Arduino platform, assisting in the work of precision beekeeping.
Fezary et al., developed a human health monitoring system, using the data of heart rate, body temperature and blood pressure as the analysis data.The collected data are processed and sent remotely through the Arduino platform to health care providers, allowing the verification and monitoring of the patient's health status.
Laskar et al., used the Arduino platform to design a meteorological system, providing information about the climate, being monitored the temperature, humidity and air pressure.
Groener et al., have described a preliminary design of a low-cost greenhouse that uses the Arduino platform and has the potential to contribute to food security in countries where the average income of the population is low.The code implemented on the platform and the greenhouse structure are designed for crops such as tomatoes, peppers and onions.
Prima et al., proposed the creation of a laboratory kit using the Arduino platform and sensors to improve quality in science education.The proposed system consists of investigating heat transfer and temperature changes at certain positions along a rod when it is being heated.
Georgieva et al., used the Arduino platform for a soil quality parameter monitoring project with wireless communication and using the concept of modular systems to carry out measurements of temperature, humidity, conductivity and soil acidity.
Gosai e Bhavsar performed the acquisition of temperature data using the Arduino platform to verify the influence of temperature on the life of a cutting tool.According to the obtained data, modifications are made in the cutting parameters, aiming at the best use of the tool.
Pocero et al., have structured a real-time monitoring of several school buildings with the objective of improving energy efficiency.The system uses the Arduino platform to monitor and manage temperature, humidity, presence, noise level and energy consumption signals.As verified in the cited examples, the Arduino platform is widely used in the most diverse types of research, aiding in the process of data acquisition and processing.
Experimental Section
The following materials were used to assemble the measuring equipment: one Arduino Mega board (Figure 1a), three K-type thermocouples (Figure 1b), three thermocouple modules (Figure 1c), a real-time clock module (Figure 1d ), a data storage module (Figure 1e), a micro SD card, a network communication module (Figure 1f), and a base board module L (Figure 1g) to interconnect all modules to the Arduino Mega board.Using the temperature signal as the starting point, the temperature measuring equipment uses type K thermocouples to acquire the temperature signals in the mold, and are then transmitted analogously to the thermocouple module.
In the thermocouple module, the signal is amplified, followed by a signal filtering process to reduce noise and interference, thus improving the quality of the measurement.In the same module is realized the cold joint compensation, as well as the conversion of the signal to the digital format, which is the format of data that digital systems can operate.As a result, the thermocouple module data is sent serially to the Arduino Mega board.
The process mentioned so far is the acquisition, processing, conversion and sending of the signal to the Arduino Mega board.However, in order for the signal to be used in the creation of graphs and spreadsheets, the signal is acquired, so the real time clock module is used to provide the precise measurement time, which keeps running even during power outages, as there is a reserve power system using battery.
Due to the speed of acquisition of the temperature signals in the mold and the measurement time information, the internal memory of the Arduino Mega board would run out quickly.Therefore, the information acquired by the temperature measuring equipment is recorded on an SD card through the data storage module, which facilitates and makes available the transfer and analysis of data in other equipment only by removing the card.
The network communication module accesses the information acquired by the temperature measurement equipment and sends the data remotely to other equipment connected in the network, such as: computer, cellular, tablet.As a design feature, considering class C IP addresses, where the first three octets are reserved for network addressing, up to 254 measurement and control equipment can be interconnected.Therefore, the possibility of forming a distributed measurement system is created.
In order to carry out the programming of the processes inherent to a temperature measurement equipment, the Arduino software [9] was used and a computer physically connected by means of a USB cable on the Arduino Mega board, which also allows the monitoring of the measurement process of the temperature signals locally.
Figure 3 shows the flowchart used in the programming logic of the temperature measurement equipment, where the internal processes performed by the Arduino Mega board together with the peripherals (modules, sensors and SD card) are available.
Results and discussion
To evaluate the performance of the temperature measurement equipment developed in this research, a comparison was made between the temperature signals acquired in relation to the commercial Agilent brand data acquisition system model 34970A (Figure 4a).However, in order to avoid the influence of the K-type thermocouple error in the process of comparing the temperature values, the same thermocouple was used for both the temperature measurement equipment constructed in this research and the Agilent commercial equipment.
Therefore, the Visomes BC200 model was used, as shown in Figure 4d, where the thermostatic bath test was performed, the K-type thermocouple being immersed in an electronically controlled heated liquid solution, thus same temperature measurement medium in the test.
To verify the temperature variation in the heated oil, a 6 1/2-digit precision Fluke Model 8846A multimeter (Figure 4b) was used in conjunction with a PT100 thermistor with 4-wire connection.In Figure 4c the test assembly is visualized and all equipment used can be checked.• Temperature measurement using Fluke's precision thermoresistance and multimeter at the beginning, middle, and end of the test to check for possible variations in the temperature of the thermostatic bath oil.
Preprints
• Individual and sequential measurement of the three channels of the temperature measurement equipment developed in this research together with the K-type thermocouple.
• Temperature measurement using Agilent commercial equipment in conjunction with Type K thermocouple.In Table 1, the resistance values of the PT100 thermoresistor measured by the Fluke multimeter can be checked and in table 2 the values in Table 1 converted to degrees Celsius.Analyzing Table 2, during the test period of the thermostatic bath, it was verified that the oil presented low temperature variation, as highlighted in yellow.The maximum variations shown are: 0.020 ° C in the 50 ° C test; 0.107 ° C in the 70 ° C and 0.082 ° C in the 90 ° C assay.Therefore, considered to be stable for the comparative test between the temperature measurement equipment developed in this research and Agilent commercial equipment.
In Table 3 the data of 570 temperature measurements performed during the thermostatic bath test were compiled, where the performance of the temperature measurement equipment developed in this research in relation to commercial Agilent equipment can be verified.As a result, with respect to the worst values recorded statistically, it can be stated with 95% confidence that the maximum variation (or amplitude) of the measured temperature in the channels is 0.0305 ° C, highlighted in yellow in Table 4. Soon, considerably lower than the value of 1.00 ° C seen in Table 3, evidencing the quality of measurement of the temperature measurement equipment developed in this research.
Regarding the average temperature value, taking the Agilent commercial measuring equipment as a reference, considering the worst values recorded, statistically, it can be stated with 95% confidence that the maximum variation of the average temperature between equipment is between the values of -1.85152 and -1.82294 ° C, with an average value of -1.83723 ° C, highlighted in green in Table 5, which corroborates the mean value calculated from Table 3, highlighted in green, of 1, 84 ° C. As a result, with respect to the worst values recorded statistically, it can be stated with 95% confidence that the maximum variation (or amplitude) of the measured temperature in the channels is 0.0305 ° C, highlighted in yellow in Table 4. Soon, considerably lower than the value of 1.00 ° C seen in Table 3, evidencing the quality of measurement of the temperature measurement equipment developed in this research.
Regarding the average temperature value, taking the Agilent commercial measuring equipment as a reference, considering the worst values recorded, statistically, it can be stated with 95% confidence that the maximum variation of the average temperature between equipment is between the values of -1.85152 and -1.82294 ° C, with an average value of - To verify the operation of the temperature measurement equipment developed in this research in the measurement of temperature signals in the mold of polymer injection machines, an experiment was carried out using the Sandretto model 50/247 polymer injection molding machine and a mold of P20 steel.
In the test, the following procedures were adopted: the machine starts the process of injection of polymeric material into the mold, generating several pieces, as shown in Figure 5a, until the system acquires thermal stability, being monitored through the thermocouples inserted in the together with Agilent's temperature acquisition system, as highlighted in red in Figure 5b.Soon after, with the system operating in a stable way, the temperature in the mold is acquired.As a result of the test, it was verified that the equipment developed in the research operated in a stable manner, as seen in the graph of Figure 6, measuring the temperature in the mold without presenting oscillations or measurement noises, however, it presented a measurement error, within the expected, as seen in Table 4 and Table 5.
Conclusion
As a result of this research, a temperature measuring equipment was developed that uses as base of operation an open and modular microcontrolled platform, where it was possible to program input and output signals through programming, thus managing the connected temperature sensors in the three channels that the measuring.
As a characteristic of the Arduino platform, due to the fact that it is modular, using auxiliary modules, it was possible to meet all the stipulated objectives for the measuring equipment, being: realization of temperature measurement with equivalent quality to commercial equipment, data storage capacity, remote thermocouple temperature verification, data storage capacity on removable digital media, and the low cost for the construction of temperature measuring equipment, in the order of $ 250.00.
In relation to the tests performed to verify the performance of the temperature measurement equipment developed in this research, different types of tests were carried out, both in the laboratory, with measured and controlled environmental conditions, as well as in the field, in the plastic materials manufacturing process.
In the test using the Sandretto brand polymer injection machine and the P20 steel mold, it was possible to verify the operation of the temperature measurement equipment in an industrial application, with the temperature measurement being carried out simultaneously in three channels of the measuring equipment, with the remote monitoring being used.As a result, a possible occurrence of noise or measurement problems was verified beyond the temperature in the mold, which could be caused by the type of environment to which the measuring equipment has undergone.However, no signs of noise or measurement inconsistencies were found in the temperature files.
As for the laboratory test, the tests were performed with the objective of verifying the errors of the temperature measurement equipment in relation to the systematic and random values of the measurements performed.As a complement, a statistical analysis of the test data was performed using the Minitab ® statistical software to calculate the mean, standard deviation and confidence interval, 95% for the measured temperature measurements.
Figure 1 .
Figure 1. Materials used in the construction and testing of measuring equipment.
Figure 3 .
Figure 3. Flow chart of the measurement system.
Figura 4 .
Figura 4. Performance verification test of the measurement equipment developed in the research.(a) Data acquisition system; (b) precision multimeter; (c) thermostatic bath test; (d) Thermostatic bath equipment.
Figure 5 .
Figure 5. (a) Parts injected into the Sandretto polymer injection molding machine; (b) Measuring mold temperature using Agilent equipment.
Table 1 .
Resistance values measured by the Fluke Precision Multimeter Model 8846A.
Table 2 .
Conversion of resistance values measured by Fluke Model 8846A Precision Multimeter in degrees Celsius.
Table 3 .
Results of the temperature measurement test in the thermostatic bath.
Table 4 .
Statistical analysis of test data for validation of temperature measuring equipmenttemperature variation in measuring channels
Table 5 .
Statistical analysis of test data for validation of temperature measuring equipmenttemperature variation in measuring channels. | 3,823.4 | 2018-03-08T00:00:00.000 | [
"Engineering",
"Computer Science"
] |
Social Drivers and Algorithmic Mechanisms on Digital Media
On digital media, algorithms that process data and recommend content have become ubiquitous. Their fast and barely regulated adoption has raised concerns about their role in well-being both at the individual and collective levels. Algorithmic mechanisms on digital media are powered by social drivers, creating a feedback loop that complicates research to disentangle the role of algorithms and already existing social phenomena. Our brief overview of the current evidence on how algorithms affect well-being, misinformation, and polarization suggests that the role of algorithms in these phenomena is far from straightforward and that substantial further empirical research is needed. Existing evidence suggests that algorithms mostly reinforce existing social drivers, a finding that stresses the importance of reflecting on algorithms in the larger societal context that encompasses individualism, populist politics, and climate change. We present concrete ideas and research questions to improve algorithms on digital platforms and to investigate their role in current problems and potential solutions. Finally, we discuss how the current shift from social media to more algorithmically curated media brings both risks and opportunities if algorithms are designed for individual and societal flourishing rather than short-term profit.
Introduction
Algorithms on digital media platforms clearly provide value, as reflected in the wealth they generate for the companies using them.They highlight relevant posts, news, people, and groups and have become necessary to reduce information overload (Narayanan, 2023b).The central role of algorithms in several types of online interaction has raised concerns that they may fuel large psychological and societal issues, specifically mentalhealth issues and political polarization.First, algorithms could contribute to increasing depression, anxiety, loneliness, body dissatisfaction, and even suicides by facilitating unhealthy social comparisons, addiction, poor sleep, cyberbullying, and harassment, especially in teenagers and girls (Ritchie, 2021;Twenge, 2020;Twenge et al., 2022).Second, they may fuel hate speech, fake news, and polarization by promoting extremist and populist content or by using algorithmic filter bubbles (Bliss et al., 2020;Lewis-Kraus, 2022).
Widespread usage of digital platforms and continuous interaction with algorithms could indeed affect individual and societal well-being in important ways (Büchi, 2021).However, direct evidence supporting these conclusions remains scarce (Bail, 2021;Ferguson, 2021;Sumpter, 2018).Researchers have investigated the potential effects of digital media and its algorithms using self-reports of social-media usage and digital traces of online behavior.Yet most existing studies cannot distinguish the effects of algorithms from the general use of digital media, social behavioral patterns, or large societal changes because their traces are intermingled in these types of data (Salganik, 2019).
We aim to illustrate how algorithmic mechanisms on digital media build on societal forces and how, in combination, they influence desirable and undesirable outcomes at the individual and collective levels.We focus on algorithms that determine how data is processed and what content is presented to users on digital media, rather than the more general concept of algorithms as a set of steps to perform a task.We describe the social drivers of online interaction and how algorithms might change these dynamics.We then summarize evidence and research gaps on social, algorithmic, and societal contributions for two sample topic areas: well-being and mental health at the individual level and polarization and misinformation at the collective level.Finally, we outline open questions and research opportunities to understand whether we can improve algorithms to contribute to human flourishing, and if so, how.
Social Drivers Underlying Individual and Group Behavior on Digital Media
Social media and its algorithms are so successful because they build on ancient human needs for connection and status (Brady et al., 2020;Meshi et al., 2015;Nadkarni & Hofmann, 2012).The twin desires to get along and to get ahead are basic human motives that were crucial for survival in our ancestral environment in social groups (Cummins, 2005;Sapolsky, 2005).Status and connection are pivotal to explaining social behavior (Abele & Wojciszke, 2013;Fiske et al., 2007;Gurtman, 2009) across many domains.Examples include face perception (Todorov et al., 2008); judgments and stereotypes (Fiske et al., 2007); relationships between individuals (Schafer & Schiller, 2018) or groups (Nadler, 2016); and cultural differences in religiosity and prosociality (Gebauer et al., 2013(Gebauer et al., , 2014)).
Connection and status motives also strongly shape social interaction and interaction with algorithms on digital media (Eslinger et al., 2021;Meshi et al., 2015).The need for connection motivates participation in the lives of friends, interest in peer groups, self-disclosure of one's own experiences, and renewal of old connections, as well as the pursuit of new connections, dating partners, and groups to join.Status motives influence how we broadcast content, present ourselves, receive social feedback, and observe and evaluate what others share (Burke et al., 2020;Meshi et al., 2015).Studies show that humans are susceptible to social feedback on digital media: Likes influence how quickly people post again (Lindström et al., 2021), whether or not they consider a post successful (Carr et al., 2018), and how happy, self-assured, and popular they feel after posting status updates (Rosenthal-von der Pütten et al., 2019;Zell & Moeller, 2018).
Algorithmic Mechanisms and Other Platform Influences
All of these social motives are also ubiquitous in offline contexts, so how do algorithms and platform features change social interaction?Algorithms constantly adapt to changes in human behavior and are updated as behavior on platforms, and societal discussion about them, evolves.Humans, in turn, strive for the attention and recognition of others to gain social status, which motivates them to reproduce the behaviors that algorithms reward.The eventually observable behavior thus results from interactive feedback loops between human behavior, algorithms, and other platform features (Narayanan, 2023b;Tsvetkova et al., 2017;Wagner et al., 2021).Algorithms are designed to optimize certain metrics, which are used to rank content in user feeds or to suggest relevant accounts.Yet these optimization metrics are usually chosen to maximize the profits of corporations and advertisers (Bak-Coleman et al., 2021;Narayanan, 2023b) rather to bring about psychological and societal benefits.
The history of the Facebook algorithm illustrates how changes in metrics can affect social behavior (Merrill & Oremus, 2021;Oremus et al., 2021;Wallaroo Media, 2022), but also how little control engineers actually have over eventual outcomes within such complex emerging feedback loops (Narayanan, 2023a).In its early days, the algorithm optimized for the number of clicks, likes, and comments and the total time spent on Facebook.As users and companies learned to game the algorithm, clickbait emerged.To counter this, Facebook started maximizing the time users spent reading or watching content in 2015, which led to more passive use, more professionally produced content, less social interaction, and less sharing of original content.Because of user complaints and decreases in interaction, Facebook adapted the algorithm to encourage more "meaningful social interactions."It boosted posts by friends and family, boosted highly commented posts, and weighted the emotional-reaction buttons much more than likes.This became problematic, as the most heavily commented posts also made people the angriest.Strongly weighting angry reactions may have favored toxic and low-quality news content.Responding to complaints, Facebook gradually reduced the angry emoji weight from five times the weight of likes in 2018 to weight 0 in 2020.
Most current digital-media algorithms strongly optimize for engagement (Narayanan, 2023b;Nikolov et al., 2019).However, social success and quality of content are only partly correlated (Salganik et al., 2006).
Optimizing for popularity even seems to lower the overall quality of content (Ciampaglia et al., 2018).Engagement metrics primarily promote content that fits immediate human social, affective, and cognitive preferences and biases rather than quality content (Menczer, 2021) or long-term goals and values (Narayanan, 2023b).For instance, users are more likely to like and share lowquality content that others have already liked (Avram et al., 2020).Popularity metrics can also be gamed with inauthentic behavior, including bots, organized trolls, and fake-account networks (Pacheco et al., 2021;Sen et al., 2018).Furthermore, the interval at which an algorithm rewards behavior influences how quickly it is repeated (Lindström et al., 2021).Especially variable and unpredictable rewards, such as those on platforms with strong virality, seem more addictive (Munger, 2020b).
Other relevant platform features beyond algorithms include the vastly enlarged scale of digital compared with offline social networks.This increases audience size and magnifies differences in the influence and social status of individual users (Bak-Coleman et al., 2021).It also creates unprecedented opportunities for building connections, earning recognition, and observing others, thereby supercharging motives of social status and connection (Bail, 2021;Bak-Coleman et al., 2021;Brady et al., 2020).This increases potentially available social feedback, which notifications, likes, shares, and comments make easily accessible, immediate, and quantifiable (Brady et al., 2020).Finally, algorithm recommendations may have changed the structure of networks, increasing the frequency of triangles (Salganik, 2019;Ugander et al., 2011) and enabling interaction between distant individuals (Bak- Coleman et al., 2021).
These conditions make status comparisons particularly likely and painful (Brady et al., 2020;Munger, 2017).In large online networks, personal information about individuals is limited, whereas information about social groups is still visible.Social groups thus become the main relationships in the network, making social identities highly salient (Brady et al., 2020).Finally, larger networks mean that one encounters a larger number and diversity of individuals and opinions than in real life (Gentzkow & Shapiro, 2011;Guess et al., 2018).Digital media thus allow people to observe many (potentially very different) others and offer people unprecedented freedom to present themselves, get feedback, and adapt; they have become a central tool people use to understand themselves, understand others, and understand which groups they themselves belong to (Bail, 2021;Brady et al., 2020).As contexts in which status and groups are highly salient, digital media have become places where different groups compete for status and in-group and out-group dynamics crucially determine behavior.
When audiences are larger and more public than private, competition between groups becomes particularly strong, as discussions between political groups show, for example, on Twitter.Similarly, YouTube's reputation for toxic comments could be linked to the extremely broad demographics of its users, leading to more conflict, and to the algorithm weighting up-votes and down-votes equally (Munn, 2020).On Facebook and especially Instagram, self-presentation is more central than group competition (Cingel et al., 2022;Midgley, 2019;Storr, 2018), leading, for example, to microcelebrities (Marwick, 2015).The TikTok algorithm guarantees a small number of views for everybody, which lowers the barriers to entry compared with the more hierarchical social networks on social media.It further makes it hard to predict which TikTok videos will go viral, which could explain long unwanted scrolling experiences, more passive watching, and less social interaction overall (Munger, 2020b).
Social Drivers and Algorithmic Mechanisms Influencing Individual Well-Being
Mainstream discourse and parts of the scientific literature often fail to distinguish between social drivers, algorithmic mechanisms, and societal context because they fail to derive causal insights from correlation, present results limited to single studies and countries, take self-reports at face value, or omit the fact that effect sizes are small (see Cavanagh, 2017;Dienlin & Johannes, 2020;Orben & Przybylski, 2019b;Ritchie, 2021;Sumpter, 2018).
Concerns are often raised about algorithms on digital media harming mental health by fueling addiction, bad sleep, and social comparison (Smyth & Murphy, 2023), or about algorithms purposefully manipulating user mood (Booth, 2014).This debate usually conflates the time spent using social media with algorithmic effects.Only one study pinpointed algorithmic effects, finding that reducing positive posts in the Facebook feed reduces the likelihood of users posting positive content by 0.1% (Kramer et al., 2014).Indirect hints that social dynamics in online media may be more harmful to mental health than algorithms come from a natural experiment-the rollout of Facebook across U.S. colleges in 2004 to 2006 (Braghieri et al., 2022).At this time, recommender algorithms still played no role on Facebook.Yet the study observed that starting to use Facebook produced a moderate effect on depression and a small effect on anxiety disorders but no significant effect on eating disorders, suicidal thoughts, or attempts.Further results hinted that the negative effects arose from unhealthy social comparisons.
Other studies on short-or long-term well-being and mental health addressed only algorithmic effects as part of social-media usage as a whole.Two randomized controlled trials testing the effects of deactivating Facebook (Allcott et al., 2020;Asimovic et al., 2021) observed small to moderate decreases in anxiety, one in depression, and one in loneliness.Many other emotions did not change, consistent with an experience-sampling study testing the effects of using Twitter (de Mello et al., 2022).Life satisfaction did not change after deactivating Facebook for 1 week (Asimovic et al., 2021), but increased after 4 weeks (Allcott et al., 2020).Furthermore, specification curve analyses showed very small negative associations with social-media usage in adolescents (Orben et al., 2019;Orben & Przybylski, 2019a).
Overall, the debate about social media and individual well-being requires more nuance.Evidence for algorithms driving or reinforcing unhealthy dynamics is very thin, supporting, at best, a small effect on mood.Instead, unfavorable social-status comparisons online may harm mental health.The direction of effects between socialmedia usage and mental health is unclear (Ferguson, 2021;Luhmann et al., 2022;Orben et al., 2019); the direction and size of effects depend on who uses social media and in what way (Büchi, 2021).For instance, teenage girls or already socially disadvantaged individuals may be particularly vulnerable (Allcott et al., 2020;Heffer et al., 2019;Midgley, 2019;Orben et al., 2019Orben et al., , 2022)).Passive, extreme, or low use is related to poorer wellbeing, whereas active, social, and moderate use correlates with better well-being (Dienlin & Johannes, 2020).Furthermore, self-reports of usage and addiction do not reliably measure actual usage and tend to systematically overestimate them, more so in some users (such as girls) than others (Boyle et al., 2022;Mahalingham et al., 2022;Scharkow, 2016;Shaw et al., 2020).
Yet algorithmic effects could also emerge as slow trends or at higher levels of the complex system involving digital media and the offline world; studies on individuals in limited time periods cannot capture such effects.For example, the constant opportunity to express oneself could slowly affect independent emotion-regulation abilities, and algorithmic reinforcement of emotional content could change norms of emotional self-disclosure in relationships over time.In any case, potential risks to the well-being and mental health of vulnerable groups need to be taken seriously, and large corporations should be held responsible for preventing harm.
None of the cited studies says anything about the societal context, including achievement pressure and individualization increasing with neoliberalism (Levitz, 2023;Storr, 2018); increasing economic inequality and insecurity (Wilkinson & Pickett, 2017); more single households in wealthy societies, driving loneliness; sleep irregularities and addiction, especially in younger adults (Cocco, 2022); or general uncertainties about the future (e.g., climate change; Ingle & Mikulewicz, 2020).All of these societal developments could crucially affect mental health, with algorithms reinforcing existing dynamics but not being the primary cause.
When problems have strong social or societal root causes, solutions will require difficult political, institutional, and economic changes.To address the actual causes, we need research that disentangles which, if any, of the issues currently blamed on algorithms are driven by social dynamics or societal context.The influence of societal context is particularly difficult to pin down.It would require large-scale and longitudinal studies tracing and separating multiple interacting factors and their online and offline effects over time, including algorithmic and societal changes across platforms, nations, and cultures.Data for such studies are currently not available, but the European Digital Services Act may be a step forward (Turillazzi et al., 2023).
Social Drivers and Algorithmic Mechanisms Influencing the Collective Dynamics of Political Polarization and Misinformation
Could algorithms foster echo chambers of like-minded people and polarization (Garimella et al., 2017)?Do engagement metrics promote hate speech, radicalized content, and fake news?We highlight a few studies that help dissociate algorithmic mechanisms from social drivers.For more details, see Van Bavel et al. (2021), Ferguson (2021), andLorenz-Spreen et al. (2022).
Online echo chambers might have a more minor role than has been commonly assumed (Bakshy et al., 2015;Bruns, 2021;Guess et al., 2018;Sumpter, 2018;Törnberg, 2022) and are smaller than offline echo chambers (Gentzkow & Shapiro, 2011).Weaker online echo chambers mean that people are exposed to more people they disagree with.Similarly, digital media may increase perceived rather than actual polarization (Bail, 2021).Supporting this, a field experiment on U.S. Twitter observed increased political polarization after exposure to posts from opinion leaders of the opposing party (Bail et al., 2018) and experience sampling reveals consistent results (de Mello et al., 2022).
Increases in actual polarization are less bad than commonly assumed; there is still overlap for substantial issues in the views of political parties (Bail, 2021).Because misinformation is largely a symptom of polarization (Altay, 2022;Osmundsen et al., 2021;Petersen et al., 2022), exposure to online misinformation might also have been overestimated.Misinformation accounts for a small proportion of digital-news consumption (Altay, Nielsen, & Fletcher, 2022) and is mostly shared by a tiny minority of users (Grinberg et al., 2019;Osmundsen et al., 2021).Additionally, misinformation has been shown not to easily change beliefs or political voting behavior (Bail et al., 2020;Guess et al., 2020).
Regarding specific algorithm effects, a study on Facebook data in 2014 (Bakshy et al., 2015) found that users' social networks determined posts in their feeds much more strongly than the algorithm.Similarly, users actively engage with more partisan news than suggested by the Google search algorithm (Robertson et al., 2023).The YouTube algorithm also does not seem to radicalize many users: Only 1 out of 100,000 who started viewing moderate content later moved to farright content (Ribeiro et al., 2021).Most movement to far-right videos comes from outside the platform, and far-right videos are not more likely toward the end of sessions, where algorithmic recommendations matter most (Hosseinmardi et al., 2021).Instead, the demand for far-right content, with supply being easy, and the lack of more moderate conservative content may explain the increases in views of such content until mid-2017 (Munger & Phillips, 2020).
Overall, evidence neither shows that algorithms cause echo chambers, nor that echo chambers cause polarization.Yet algorithms can still contribute to polarization-for example, by weakening echo chambers and exposing people to more views they disagree with.Current evidence is consistent with the view that digital media as a whole, including algorithms, fuels perceived polarization by making extremist voices more visible and hiding moderate majorities (Bail, 2021).Two randomized controlled trials support this: In the politically polarized United States, affective polarization decreased after Facebook abstinence (Allcott et al., 2020).However, not having online contact with (probably moderate) ethnic out-group members in Bosnia-Herzegovina increased affective polarization (Asimovic et al., 2021).Similar to the idea of perceived polarization increasing actual polarization, the myth of fake news being common makes people more skeptical of news in general (Altay, Berriche, & Acerbi, 2022;Fletcher & Nielsen, 2019;Guess et al., 2021).Again, most studies on digitalmedia effects say little about larger societal drivers of polarization.One likely driver of increasing affective polarization, and thus misinformation, is the rise of authoritarian populism in many Western countries, which itself may arise from economic insecurities or backlash to progressive cultural change (Inglehart & Norris, 2016; but see Schäfer, 2022).
Yet such societal developments can interact with algorithmic effects by affecting discourse and decisions about algorithms.Letting platforms decide how to rank content may have seemed obvious for a long time, but discussions about this are increasing.Additionally, currently polarized or populist debates may make it difficult to find common ground on algorithmic optimization metrics, making it harder to address potentially negative effects.Similar feedback loops in the positive direction could begin with algorithms that emphasize the overlap in views of political groups, which could reduce polarization.Furthermore, algorithms that emphasize nuanced content could help decrease paralyzing climate anxieties or highlight constructive perspectives that motivate action.Finally, algorithms could create more collective emotional experiences by facilitating the spreading of emotional content.This could motivate protest movements or prosocial behavior but also foster intergroup conflict and intolerance.
Research Avenues Toward Solutions and Flourishing
Digital-media companies benefit from the narrative of omnipotent algorithms, as their business model relies on their customers (i.e., advertisers) believing it (Munger, 2020a;Sumpter, 2018).For instance, Cambridge Analytica wanted its customers to believe they could shift political opinion in the crucial target group of undecided voters (Sumpter, 2018).Munger (2020a) argued that activists and society should stop buying this story.Silicon Valley corporations should carry responsibility for evaluating the potential societal consequences of their platforms.Still, blaming technology as the supposed mechanism behind a problem without looking at the drivers that power the problem is unlikely to lead to resolution.This approach directs attention away from actual root causes and potentially misleads societal discourse and policies, creating ground for further complaints.
Famous platform critics such as Francis Haugen or Elon Musk (Oremus et al., 2021;Riemer & Peter, 2022) have suggested getting rid of algorithms entirely and returning to reverse chronological ordering of posts.However, chronological order is just another kind of algorithm with its own drawbacks (Riemer & Peter, 2022): It favors more frequent posters, does not reduce information overload, and likely implies that users will miss more carefully prepared but rarer content.Getting rid of algorithms also means not using them as tools where they are indeed useful.Using algorithms well, in turn, requires developing shared visions and valuesthings users want algorithms to align with-which is a major important avenue for future research.
Future researchers need to develop and test theories about the role of algorithms (see Box 1), including potentially positive contributions and the mechanisms and outcomes of feedback loops with social behavior.Algorithms could even help to solve problems to which they currently contribute, and they can be intentionally designed to foster short-and long-term well-being and flourishing (Steinert & Dennis, 2022).This requires developing a vision for digital-media design and algorithm design beyond those proposed by existing forprofit companies (Bail, 2021;Büchi, 2021).
Although problem audits of algorithms are rare, studies on beneficial effects are even rarer.Some A/B tests on beneficial outcomes exist for interface design (e.g., Zhang et al., 2022), content manipulations and connection recommendations (e.g., Rajkumar et al., 2022), or for achieving collective outcomes with the help of random bots (Shirado & Christakis, 2017).However, more experiments comparing different optimization algorithms and comparing platforms with and without algorithms are needed.
Testing the effects of current and possible future algorithm and platform design requires platforms that allow experimental manipulation while obtaining users' consent.Computational social scientists have begun developing such bespoke social-media platforms to test the effect of concealing political affiliation or gender identity (Combs, Tierney, Alqabandi, et al., 2022; Box 1.Some Important Psychological Research Questions on Algorithms on Digital Media
Overarching questions
• With which values and purposes do we want the outcomes of our algorithms to align?
• How can psychological knowledge about social behavior and cognition help to design algorithms and platforms to best foster human well-being and flourishing at an individual and collective level?• Do current algorithms on digital media have beneficial effects compared with media without algorithms?
• Which new risks and opportunities arise from the current shift from social to algorithmic media?• Which platform design and algorithm features best align with different purposes?Which purposes require different digital environments, and which can be combined?• How does giving users choices about algorithm metrics or other design features affect individual and collective well-being?Which choices should be made available, and which should be implemented broadly?How should these choices be assessed?
Individual and interindividual well-being, happiness, and flourishing • What are specific algorithmic effects on emotions and mental health, independent from social-media usage as a whole?• How could algorithms and other platform design features . . . foster short-or long-term well-being and flourishing of individuals (including pleasure, happiness, life satisfaction, and connection)?Combs, Tierney, Guay, et al., 2022), social-engagement metrics (Avram et al., 2020), or anti-addictive design features (Zhang et al., 2022).Collaborations between academia and existing platforms are another promising approach (Stray & Hadfield, 2023).
Ideas for algorithm and platform design to foster flourishing
Algorithms can improve digital-media platforms in two ways: by using different optimization metrics to rank content, or by prompting interventions upon detecting problematic content.Current digital-media platforms show that engagement metrics that optimize for entertainment are unlikely to foster rational debates.Optimal design choices will thus likely depend on the purpose of a platform and potentially on user preferences.We may want to create different platforms to foster nuanced political discussion, amplify entertainment and shortterm pleasure, promote regular contact between friends and relatives, deepen personal relationships, or build communities (e.g., for mental-health support).Future research on which design choices work best to achieve each purpose, and which ones require separate platforms or subspaces on existing platforms, would be very valuable.Platforms for fostering nuanced political discussions that strengthen social cohesion, moderate voices, and diversity will have to focus on reducing perceived polarization.This requires reducing the visibility of strongly partisan and triggering content (Rose-Stockwell, 2018), perhaps with algorithm metrics that prioritize content popular on both sides of the political spectrum.This would highlight moderate voices and reveal the opinion overlap for the important issues where it actually exists (Bail, 2021), and could promote more trustworthy news sources (Bhadani et al., 2022).Similar algorithms could highlight which principles or practical approaches resonate with people on both sides of other belief spectrums, such as those relating to climate change or alternative medicine.
Algorithm rankings could further foster intergroup contact and understanding by presenting posts that are not too distant from a user's own position (Levendusky, 2018;Sherif, 1963).In this way, algorithms could support small steps toward understanding alternative views.Using algorithmic estimates of users' positions on a dimension, platforms could further label extreme voices as such, give users feedback about their own position, or show how moderate and extreme users on both sides have responded to an account or post (Bail, 2021).Other suggestions include toning down status incentives by hiding or reducing the visibility of engagement metrics for certain types of posts (Avram et al., 2020) or adding cues that spotlight passive user behaviors (e.g., how many scrolled over a post; Lorenz-Spreen et al., 2020).Twitter introduced view counts for tweets early in 2023, creating research opportunities to explore how this affects social-reward experiences or the spreading of polarizing and untrustworthy content.Finally, anonymity is a promising nonalgorithmic design feature for reducing conflicts rooted in social identity, with the potential to make discussions on controversial issues kinder (Combs, Tierney, Guay, et al., 2022).
Optimizing algorithms for metrics such as civility (Lewandowsky & Kozyreva, 2022;Oremus et al., 2021) would require defining what counts as civil and how civility fosters democratic discourse and diversity.When a minority is unjustly neglected, or an elite unfairly privileged, for example, angry responses are appropriate and necessary.Rather than deciding which values should guide the choice of algorithm metrics, platforms could also let users define values themselves (Lewandowsky & Kozyreva, 2022;Lorenz-Spreen et al., 2020).Facebook has tested such an approach in its "breaking the glass" experiments, deploying an algorithm that emphasized posts that users considered to be "good for the world" (Roose et al., 2021).Although this reduced low-quality content, it also lowered how often users opened Facebook and was therefore implemented only in a weakened version.
A second way to use algorithms is to detect certain posts or activities and then trigger interventions.The simplest of all interventions is adding friction, that is, increasing the time or effort it takes to share content (Brady et al., 2020;Lorenz-Spreen et al., 2020;Menczer, 2021).Adding friction seems particularly useful to prevent impulsive sharing of sensational news and outraged or toxic comments.In some cases, simply adding a time gap before allowing users to post or share might suffice.In others, additional prompts could encourage reflection before sharing (Rose-Stockwell, 2018).Empathic and humanizing prompts have been shown to reduce affective polarization (Saveski et al., 2022) and racist harassment (Hangartner et al., 2021;Munger, 2017).Undo prompts after posting hateful comments, default options to turn comments from public to private, or ideological prompts explaining that posts with moral-emotional language are unlikely to reach the other side, could all reduce hateful content (Rose-Stockwell, 2018).Interventions that effectively reduce affective polarization provide further inspiration (Hartman et al., 2022;Voelkel et al., 2022).
To foster mental health and healthy usage of digital media, algorithms can detect linguistic markers of symptoms or certain activity patterns.Trying to detect users at risk of mental-health issues, with the goal of then providing contact points for support, is a popular research field, which, however, urgently requires methodological-validation efforts (Chancellor & De Choudhury, 2020).To reduce unwanted addictive use, algorithms can encourage users to disengage by providing reminder bots after excessive periods of scrolling or providing usage statistics (Zhang et al., 2022).Interventions such as reading-progress indicators, feed filters and content blockers for specific types of content, and separate topic-focused feeds instead of one main feed, seem even more effective (Zhang et al., 2022), and could be improved via algorithmic suggestions.Finally, we do not know of any research on how changing algorithm metrics could support individual well-being.Algorithms that reduce the visibility of toxic, regrettable, and outraged content may help reduce content that negatively affects well-being (Rose-Stockwell, 2018).Research on algorithms that prioritize content from important personal contacts, expressions of empathy and connection, or prosocial behaviors could contribute to positive well-being outcomes.
Choosing values, validating metrics, and evaluating their effect on outcomes
As the above discussion illustrates, many different values are potentially justifiable candidates for algorithmic optimization.Choosing such values, validating the metrics to optimize for them, and testing their effect on various outcomes will require research in cultural, moral, social, political, affective, and clinical psychology, as well as computational, sociological, political, and economic approaches.Such research needs to determine for which values societal consensus is possible, and where digital media have to accommodate different needs, visions, and goals within the same or between different platforms.It should further explore which values and goals individuals prioritize and how social and cultural norms affect these processes in communities and societies.
Research could also compare different ways of assessing these value preferences: Avoiding perpetuating the influence of social and cognitive biases will probably require asking for user decisions in advance and at an abstract level, rather than measuring immediate preferences when users thoughtlessly scroll through their feeds.Preliminary research shows that most U.S. users across political and demographic groups opt for seeing more accurate, nuanced, friendly, positive, and educational content (Rathje et al., 2022), although such content currently does not typically go viral by itself.Researchers need to test whether users would actually make the choices they report preferring on the platforms they regularly use, and they then need to determine whether this would reduce misinformation and polarization at the macroscale of digital platforms.They further need to explore where individual user or community choices on algorithmic rankings or interventions are possible and beneficial (Lewandowsky & Kozyreva, 2022;Lorenz-Spreen et al., 2020), and where they need to be restricted.For example, letting users opt for only partisan content is dangerous, as contact with moderate voices from other groups may be necessary to reduce polarization.
Once certain values are agreed upon, methodological research can be employed to validate which metrics could actually represent those values, relying on the digital trace data available to algorithms.Finally, empirical research should be used to investigate how different metrics would affect the various outcomes, including affective well-being, mental health, societal cohesion, and nuanced political discussions.Given that social media are complex systems with emerging feedback loops between social drivers and algorithms, this research needs to incorporate methods from complexity science and computational social science, such as network analysis or agent-based modeling (Borsboom et al., 2021;Jackson et al., 2017;Smith & Conrey, 2007;Vlasceanu et al., 2018) to address these many open questions.
Shift from social to algorithmic media
Twitter, Facebook, and Instagram could be referred to as traditional social media, as their information-distribution mechanism relies primarily on social networks (Mignano, 2022;Munger, 2020b).In contrast, other platforms like YouTube and TikTok mostly rely on recommendation algorithms instead of social links.On traditional social media, social drivers can have a much larger influence on interactions and spreading dynamics.In contrast, on algorithmic media, the platform itself has much more power to determine presented content through the recommender system and content feed (Mignano, 2022;Narayanan, 2023b).Algorithms are more economically competitive as information-distribution mechanisms because social graphs are now easily available (Mignano, 2022).Likely for this reason, Facebook and Instagram have started following TikTok's example by adding short recommendation-based video feeds.This trend may entirely change our current conclusions about the algorithmic effects that have been limited so far.Algorithmic media might worsen problems like addiction or propaganda.Munger (2020b) has argued that the immersive nature of TikTok's mobilefirst design, its higher capacity to evoke emotions via both visual and audio information, the ease of posting content, and the unpredictable virality of its algorithm might make it more addictive and its users more vulnerable to political persuasion.However, the shift from social to algorithmic media may also present an opportunity for the endeavor of designing digital media that foster human flourishing.Because algorithmic content distribution gives greater control to the platform compared with popular users, algorithms could select content on the basis of metrics that foster well-being of individuals and societies.They could highlight overlap between opposed groups (Bail, 2021), prioritize news a user actually wants to see (Rathje et al., 2022), or simply limit how far fake news spreads (Bak-Coleman et al., 2022).Both new risks and opportunities arising from algorithmic media are important avenues for future research in psychology and computational social science.
If algorithms make societally relevant decisions, it becomes pivotal who takes these decisions, and in what way.Making sure these decisions benefit society will require transparency about algorithmic design (Kozyreva et al., 2021;Wagner et al., 2021).The recent release of the code of the Twitter algorithm illustrates that in order to actually evaluate effects we need not only information about how the algorithms weigh types of content and interactions but also information about the machine-learning models that make suggestions for individual users (Narayanan, 2023a).Although we see potential in beneficially using algorithms on digital media, we must acknowledge the barriers that exist for this kind of research.Because the vast majority of online media are proprietary for-profit platforms, the designs and targets we presented are likely at odds with profit-making to a certain extent.Testing, implementing, and adopting solutions will therefore likely require regulation (Gal, 2022).Given the unique role of digital media in creating a public sphere in a globalized world, researchers and activists have even discussed whether digital media should become a public good (Fournier-Tombs, 2022).However, we are also still very early in the history of digital-media platforms, with large shifts of users to new platforms every couple of years (Bail, 2021).Over time, market dynamics could still play out in ways that better satisfy user preferences beyond short-term rewards.
Conclusion
We have outlined different ways in which algorithms on digital media could promote positive emotions, mental health, social cohesion, and nuanced discourse.In the context of a globalized world, polarized democracies, and increasingly individualized societies, efforts to design algorithms that foster intergroup contact via digital media may make valuable contributions to reduce social, ethnic, political, and cultural barriers.
Transparency Action Editor: Melanie Mitchell Editor: Interim Editorial Panel Declaration of Conflicting Interests
The author(s) declared that there were no conflicts of interest with respect to the authorship or the publication of this article.
emphasize connection over social comparison between individuals? foster new and deepen existing important relationships? reward interindividual empathy, support, and prosocial behavior? foster successful emotion regulation or collective emotional experiences?• How would allowing users to choose what they want to see affect their well-being?Social cohesion, nuanced political discussion, and high-quality information • Do algorithms contribute to increasing affective polarization by fueling perceived polarization-for example, by suggesting more extreme political content? • How could algorithms and platforms . . . make silent moderate majorities more visible? reduce the visibility of extremist, toxic, outraging, or regrettable content? promote nuanced and high-quality content? reward expressions of empathy and understanding? reward cooperation rather than status competition between groups? promote constructive online intergroup contact?• How would allowing individual user or community choices on algorithmic ranking affect polarization and the spread of misinformation at the macroscale on digital platforms? | 8,237.6 | 2023-07-19T00:00:00.000 | [
"Computer Science",
"Sociology",
"Psychology"
] |
A Recommender System for Mobility-as-a-Service Plans Selection
: Transportation and mobility in smart cities are undergoing a grave transformation as new ways of mobility are introduced to facilitate seamless traveling, addressing travelers’ needs in a personalized manner. A novel concept that has been recently introduced is Mobility-as-a-Service (MaaS), where mobility services are bundled in MaaS Plans and offered to end-users through a single digital platform. The present paper introduces a recommender system for MaaS Plans selection that supports travelers to select bundles of mobility services that fit their everyday transportation needs. The recommender filters out unsuitable plans and then ranks the remaining ones on the basis of their similarity to the users’ characteristics, habits and preferences. The recommendation approach is based on Constraint Satisfaction Problem (CSP) formalisms combined with cosine similarity techniques. The proposed method was evaluated in experimental settings and was further embedded in real-life pilot MaaS applications. The experimental results showed that the proposed approach provides lists of MaaS PlanMaaS Plans that users would choose in a real-life MaaS setting, in most of the cases. Moreover, the results of the real-life pilots showed that the majority of the participants chose an actual MaaS Plan from the top three places of the recommendation lists.
Introduction
Mobility-as-a-Service (MaaS) is a novel concept for smart mobility ecosystems that aims to integrate the various transport services available within a city/area, accessible on demand and through one single digital platform, according to Datson [1] and Moura [2]. A fundamental objective of MaaS is to serve the transportation needs of people by providing personalized mobility solutions in tailored bundles adjusted to user's requirements, as per Kamargianni et al. [3]. Recent investigations by Sochor et al. [4] and Robinson [5] show that MaaS can offer "something to everyone", highlighting its social inclusion character. From this perspective, it is commonly believed that MaaS will enhance the traveling experience, diminish travelers' expenditure and effectively coordinate travel demand, albeit refining the environmental and social impact of mobility in accordance with Durand [6]. As reported by Arias-Molinares & García-Palomares [7], there is an increasing number of MaaS concentrated scientific articles released through years, yet there are principal queries to be resolved, in the path of implementing its visions and sustainable character.
MaaS has emerged as a result of profound transformations and disruptions in smart cities' transport systems driven by mobility and technological innovations. Transport options, which until recently commonly relied on public means (bus, metro, rail) and the private car, are now being expanded with the emergence of new mobility solutions that are characterized by greater flexibility and take advantage of sharing economy concepts (e.g., bicycles, electric scooters, car sharing, etc.), of which travelers aspire to be a part of, by exploiting the advantages of sharing their mobility assets [2]. As Sakai [8] distinctly claims, MaaS can be reflected as all the modes of transport other than private automobiles. Such solutions lead to optimized use of resources and contribute to sustainability targets with lower greenhouse gas (GHG) emissions, in line with Eliot et al. [9] and Cruz and Sarmento [10], and energy consumption, following Becker et al. [11], while allowing travelers to reach their destination with different options or by even combining different options in what is commonly termed multimodal mobility. In terms of technological advancements, the advent of intelligent and integrated transport systems as well as the ubiquitous use of smartphones and related transport applications provide the means to address the inherent complexities of modes' integration and related challenges with respect to seamless payment, booking and journey planning, as well as sharing of revenue between the distinct modes and operators, while the competitiveness among the various engaged operators may contribute to upgraded services [2]. Besides, Kamargianni et al. [12] argue that higher mobility integration is preferable by the travelers, whereas user-friendly MaaS apps could increase users' aspiration to get involved with this innovative scheme.
MaaS is provided by MaaS Operators, i.e., transport companies that are engaged with the duty of mediating and entering into agreements with public and private mobility operators on a city, intercity or national level in order to provide bundled transport services, referred as "MaaS Plans" or packages or mobility products [6]. The MaaS Operator incorporates the various mobility service providers' offerings, creates the MaaS products and merchandises them to end-users based on Kamargianni et al. [13]. MaaS operators deploy specially designed mobility apps and relevant back-end platforms that offer to travelers a single point for MaaS Plans selection, route planning, booking and payment, while recent work of Esztergár-Kiss et al. [14] conveys the continuous growth of the MaaS market in multiple layers from new geographical deployment areas to new business models.
In a MaaS context, there can be a variety of MaaS PlanMaaS Plans with diverse transport services levels, which are designed to address the needs of dissimilar categories of travelers within a specific area (e.g., a MaaS PlanMaaS Plan may consist of a monthly public transport pass, a number of taxi and bike sharing rides). Essentially, MaaS PlanMaaS Plans are mixtures of the transport services offered by mobility operators of a city/area and its distinct characteristics with whom MaaS operators have engaged into agreements based on Esztergár-Kiss and Kerényi [15]. Instances of relevant transport services aggregated within MaaS PlanMaaS Plans may be public transport, taxi, car sharing, bike sharing, car rental, ride-hailing, e-scooters and/or other related services such as parking or e-vehicle charging stations. It is evident that the set of MaaS Plans for a particular city constitutes a selection space that grows depending on the available transport services, the combinations of which can produce large choice assortments with compound configurations, a setting that is often present in numerous e-commerce cases and is tackled by the theory of recommender systems (RS) that copes well with such information excess challenges as said by Ricci et al. [16].
The first step that travelers need to follow when enrolling in MaaS is to identify and buy a MaaS PlanMaaS Plan that fits their transport needs and desires effectively and adequately. However, travelers are commonly accustomed to using single transport services, and in the newly introduced concept of MaaS that delivers bundled mobility services, the cognitive assignment of identifying the proper MaaS PlanMaaS Plan can be cumbersome, may not be easily handled and can hinder the widespread adoption of MaaS. As Felfernig et al. [17] observe, the task of addressing the best-matching products is a compound decision-making process due to the users' limited awareness of the domain and calculation efficiency. Under this prism, tools and mechanisms are needed that facilitate end-users to select MaaS PlanMaaS Plans by identifying and recommending those that are relevant to users' specific characteristics, habits and preferences.
In this paper, we present a hybrid knowledge-based recommender system that supports MaaS users to identify plans suitable for their needs and preferences. The system relies on Constraint Satisfaction Programming (CSP) theory to filter the initial search space of available MaaS Plans into a subset that matches a specific user's profile and is coupled with a similarity calculation mechanism that considers users' mobility habits and the filtered MaaS PlanMaaS Plans in order to deliver a ranked list of MaaS PlanMaaS Plans in which the plans that are more similar to the user are placed in top positions. The proposed approach was tested and compared against different variations and recommendation strategies in a controlled experiment in which 262 users participated in a within-subjects study design: they were presented with two lists of MaaS PlanMaaS Plans, each one generated by a different variation, and were asked to evaluate them. Moreover, the approach was implemented as part of a recommender system service that was integrated in a fully fledged MaaS app as part of the MaaS4EU research project (http://www.maas4eu.eu/ accessed on 21 July 2021). The MaaS4EU app provided the means to evaluate the proposed approach in real-life situations where users from the pilot cities of Budapest (Hungary), Luxemburg and Greater Manchester (UK) selected, bought and used MaaS PlanMaaS Plans for their everyday transportation needs. The results indicate that the MaaS PlanMaaS Plans recommendation lists generated by the proposed approach are preferred compared to other recommendation mechanisms as well as that the proposed system is able to suggest MaaS PlanMaaS Plans that fit user preferences and needs, as the plans users selected in the real-life applications were within the top three positions in the majority of the cases.
The remainder of the paper proceeds as follows: Section 2 presents an overview of the related work, while Section 3 provides a detailed description of the proposed recommender system. Section 4 focuses on the design, deployment and results of the controlled experiment, and Section 5 demonstrates the real-life evaluation that was performed in the context of the MaaS4EU pilots. The paper concludes in Section 6 with a discussion of our results, future research directions and final remarks.
Related Work
Recommender systems (RS) filter out information and suggest items of interest to users based on their preferences. In many instances, it is required to make choices with insufficient personal knowledge of the available selections as specified by Resnick, P., and Varian [18]. In the present era of information overload, Ricci et al. [16], state that recommendation technologies are being used in many application domains including news, music, e-commerce, movies, etc. A recommendation system relies on methods that model user preferences, which can be captured either explicitly or implicitly, and item characteristics. Recommendation algorithms process the available user and item information and try to identify possible connections between items and users in order to suggest the most relevant items, maximizing the value of the matching between a specific item and a specific user, according to Lops et al. [19]. In addition, Konstan and Riedl [20] convey that a strong point of RS is that they moderate the working load of users, who are deluged by the potential number of alternatives. By approaching the problem of introducing users to the novel domain of MaaS with the goal of recommending them a MaaS package that corresponds to their needs, the RS theory and its successful applications among other domains appears to be a rather ideal solution.
According to Aggarwal [21], the underlying idea of RS is that certain dependencies occur between user-and item-centric actions. In RS theory, there is a long catalog of various learning models used to complete the task of inferring users' interests around the available items/products. Indicatively, the "collaborative filtering" class concerns the use of ratings collected by several users in a collaborative form to predict absent ratings. "Content-based" RS are grounded on the principle that user interests can be modeled by means of properties of the items they have either rated or accessed in the past. An alternative family of RS is that of "knowledge-based" systems, where users define interactively their interests, and the user requirements are mixed with domain knowledge to deliver recommendations.
There are yet advanced models, where contextual information such as temporal data, external know-how, location, social or network information may be utilized to provide recommendations. Data-driven collaborative filtering and content-based RS require past ratings or contextual data that, in the present novel domain of MaaS, are not available; thus, our work focuses on knowledge-based recommenders, including RS for bundles of product or service concepts detailed in the following.
In the initial setup and deployment of a RS, an important issue arises in the absence of past historical data and user input, the so-called cold-start problem. As Felfernig et al. [22] indicate, a type of RS that efficiently addresses this kind of challenge by exploiting explicitly stated user preferences and knowledge of the under-investigation field is the knowledgebased recommender systems. Related approaches explicitly ask users to provide their preferences and needs, which are recorded in a user profile, while knowledge engineers codify the domain experts' knowledge into a proper and runnable representation; thus, the system generates suggestions based on this information. Our work relies on constraintbased approaches that are grounded on the constraint programming theory. Constraintbased recommenders principally utilize predetermined recommender knowledge bases that incorporate explicit rules about the way that users' needs are related with item properties. More specifically, a Constraint Satisfaction Problem (CSP) is considered as a task that entails detecting a value for each variable included in a predefined set of variables, where constraints indicate that some subsets of values cannot be used in tandem, according to Freuder and Mackworth [23].
Constraint-based mechanisms have been extensively utilized within recommender systems in a range of business fields. A domain-independent knowledge-based recommender system, presented by Felfernig et al. [24] and named CWAdvisor, assists in the process of product selection through a personalized conversation with successful applications in financial services and the electric goods market. Another system is described by Jannach et al. [25] under the name of "VIBE", which essentially represents a virtual advisor supporting tourists in their decisions. "VIBE" grounds its implementation on a knowledge-based conversational recommender that delivers personalized plans offered by a spa resort and are adjusted with their potential customers' preferences. The work demonstrated by Reiterer et al. [26] describes a constraint-based recommender that supports households in selecting the optimal waste disposal strategy. Equivalently, Murphy et al. [27] propose an energy-saving recommender system, which makes use of real-world energy consumption data of appliances and delivers behavioral change suggestions along with an optimal appliance schedule recommendation, with the goal for the users to achieve their energy saving goals. Furthermore, Zanker et al. [28] use constraint-based modeling to approach the complex task of composing product bundles, in particular, travel bundles that collaborate accommodation with activity choices through a generic purpose platform called web configurator. The configurator integrates recommendation mechanisms with CSP principles and delivers a set of personalized product bundles, adjusted to tourists' requirements while following the e-tourism field constraints, through a hybrid approach that merges collaborative filtering with knowledge-based techniques.
Producing recommendations and personalized assortments of bundles for products or services is a research question that has been studied in domains such as tourism, telecommunications and e-commerce. A review of the frequently used types of RS that solve the problem of real-time touristic services' (e.g., activities, places to stay) configuration by dynamic packaging is presented by Schumacher and Rey [29], where they present as most useful RS, the association rules, the conversational and preference-based RS. The work of Beheshtian-Ardakani et al. [30] focus on the challenge of suggesting product bundles for e-commerce platforms from a marketing point of view and propose a new model that uses market segmentation variables along with customer loyalty analysis. In particular, the product bundles are specified for each market segment by clustering algorithms and association rules, while further for the recommending task, classification models are utilized. Zhang et al. [31] propose a hybrid recommender system, applied in the telecom domain, that incorporates user and item collaborative filtering techniques with a set of fuzzy methods and knowledge-based rules, tackling the problem of vast assortments of services/products that are available and customers are requested to choose from. More recently, Kouki et al. [32] suggest the use of product hierarchies with transactional or domain knowledge data, leading to possible compilations of product assortments. For the recommendations' generation, a deep similarity model that exploits the textual embedding is constructed using Long Short-Term Memory (LSTM) networks and is further evaluated for a big online retailer. In addition, Dragone et al. [33] suggest a machine learning recommender system applicable in the telecommunication and multimedia domains that exploits the constructive preference elicitation framework with coactive learning, called Smart Plan. Bai et al. [34] approach the personalized bundle list recommendation challenge as a structured prediction problem introducing the bundle generation network by utilizing encoder-decoder architecture with a range of techniques that ensures the good quality and heterogeneous bundle list with a proper package size.
Tackling the MaaS Plans selection problem, given its significant drawback of past data absence of both users and MaaS packages motivated us in applying recommendation technologies for providing a personalized plans selection process experience in MaaS settings, guiding the travelers through this novel framework of MaaS and eventually recommending MaaS Plans to them. A main contribution of our work is the configuration and application of the selected RS methods to the problem of personalizing the MaaS Plans selection process, as described in Section 3. We have defined a novel approach toward capturing user preferences, subsequently eliminating the search space from the plans that are not in line with the stated preferences of the user and eventually deriving the similarity among the remaining plans and the user's shaped profile. To our knowledge, the pertinent problem of personalizing the MaaS Plans selection process has not been addressed in prior literature, albeit RS approaches could certainly be proved as a useful tool in solving it.
This work builds on top of our approach previously described in [35], where we introduced the use of constraint models for MaaS Plans recommendations and set the ground for tackling the MaaS Plans selection problem. In this work, we have introduced improvements in the constraint models and the employed similarity metric mechanism, including a data-driven mechanism aiming to exploit past users' data; we have implemented and integrated the approach in a real MaaS application; we have evaluated the approach in both experimental and real settings; and we have analyzed the results of the evaluation, which provide useful insights for both research and practice. Figure 1 depicts the place of the proposed MaaS Plans Recommender in a MaaS framework. The recommender has been developed to become a useful tool for MaaS endusers that will be benefited by identifying bundled mobility solutions, in accordance with their individual habits and needs, delivered by the system. The system offers the required operations toward capturing user preferences, subsequently eliminating the search space from the plans that are not in line with the stated preferences of the user and eventually deriving the similarity among the remaining plans and the user's shaped profile. The final result is a filtered and ranked list of MaaS Plans from which the user is able to choose the plan that better conforms with her/his needs.
Approach
More details of the proposed MaaS Plans Recommender are shown in Figure 2. The recommender operates on a set of MaaS Plans which contain varying mode levels depending on the available transport modes and agreements of the MaaS operator with the individual service providers. The MaaS Plans are, in general, received by transport engineers in Excel/CSV format and are further processed by transformation scripts in order to be transformed into JSON format and subsequently be used by the recommender system. It should be noted that the various levels of the transport modes included within the MaaS Plans are provided by the MaaS Operator. The MaaS Plans Recommender utilizes constraint-based filtering that leverages a prespecified knowledge base, including explicit rules (constraints) regarding how to relate user characteristics and habits to the existing MaaS product attributes (cf. Figure 2 "CSP-based filtering"). Moreover, a similarity function infers the proximity between the user and the filtered MaaS Plans list in order to generate a ranked list of MaaS Plans to be recommended to end-users (cf. Figure 2 "Similarity-based Plan Ranking"). The similarity mechanism receives direct user feedback in terms of stated users' transportation habits and willingness to include different mobility services in their MaaS Plan and, additionally, past data (i.e., past subscriptions) when available. With respect to past data, these can be captured for users who have already subscribed to at least one MaaS Plan and have made use of it; thus, the system exploits this information with the goal to further enhance the results it delivers. More details of the proposed MaaS Plans Recommender are shown in Figure 2. The recommender operates on a set of MaaS Plans which contain varying mode levels depending on the available transport modes and agreements of the MaaS operator with the individual service providers. The MaaS Plans are, in general, received by transport engineers in Excel/CSV format and are further processed by transformation scripts in order to be transformed into JSON format and subsequently be used by the recommender system. It should be noted that the various levels of the transport modes included within the MaaS Plans are provided by the MaaS Operator. The MaaS Plans Recommender utilizes constraint-based filtering that leverages a prespecified knowledge base, including explicit rules (constraints) regarding how to relate user characteristics and habits to the existing MaaS product attributes (cf. Figure 2 "CSP-based filtering"). Moreover, a similarity function infers the proximity between the user and the filtered MaaS Plans list in order to generate a ranked list of MaaS Plans to be recommended to end-users (cf. Figure 2 "Similarity-based Plan Ranking"). The similarity mechanism receives direct user feedback in terms of stated users' transportation habits and willingness to include different mobility services in their MaaS Plan and, additionally, past data (i.e., past subscriptions) when available. With respect to past data, these can be captured for users who have already subscribed to at least one MaaS Plan and have made use of it; thus, the system exploits this information with the goal to further enhance the results it delivers.
CSP-Based Filtering
Under the CSP basis, two discrete phases of the problem-solving process are specified: (i) the problem is modeled as a set of product and customer variables, and (ii) a set of constraints are determined and applied on these variables and should be satisfied in order to derive a solution. When the defined constraints are applied, products that do not
CSP-Based Filtering
Under the CSP basis, two discrete phases of the problem-solving process are specified: (i) the problem is modeled as a set of product and customer variables, and (ii) a set of constraints are determined and applied on these variables and should be satisfied in order to derive a solution. When the defined constraints are applied, products that do not satisfy these constraints are filtered out. The definition of the Constraint Satisfaction Problem (CSP) as well as its solution are described below: Definition (Constraint Satisfaction Problem-CSP)-A Constraint Satisfaction Problem (CSP) can be defined by a triple (Vc, Vprod, C), where Vc describes customer properties or requirements, Vprod describe the properties of a given product assortment PROD, and C represents a set of constraints that can include customer constraints restricting the possible instantiations of customer properties; product constraints defining restrictions on the possible instantiations of product variables; and filter conditions defining restrictions on the possible combinations of customer properties and product properties.
Definition (CSP Solution)-A solution for a given CSP = (Vc, Vprod, C) is represented by a complete assignment to the variables of (Vc, Vprod) such that it is consistent with the constraints in C and results in a set S of prod i ∈ PROD.
The CSP approach could potentially cater to an arbitrary number of product and customer properties as well as constraints. In this work, the customer properties refer to user transportation habits, which are explicitly provided by the user through a MaaS app for the modes included within the available MaaS Plans as provided by the MaaS Operator. For each available mode, users state how often they use it in a four-level scale: (i) Never, (ii) Once/few times per month, (iii) Once/few times per week, (iv) Every day. Moreover, we consider the availability of a driving license as an extra property, which affects modes related to the use of a car such as car sharing and car rental. In terms of product properties, these include the mode levels in the MaaS Plans and related mode types (e.g., car sharing, public transport etc.).
The constraints are divided in two separate groups: the hard and the soft constraints. Hard constraints filter out MaaS Plans that contain specific modes or MaaS Plans that do not contain a specific mode, whereas soft constraints filter out MaaS Plans that contain specific mode levels.
The two principal classes of hard constraints (HC 1 . . . n ) are as follows: • HC 1 , "If user model.driving license = 'No', CarSharing = 0", meaning that, in case a user does not possess any driving license, MaaS Plans including car sharing are filtered out. • HC 2 , "If user model.mode_i_usage = 'Every Day', Mode_i ! = 0", indicating that frequent users of a specific mode will be delivered MaaS Plans that definitely include this mode and exclude the ones that do not include it, with the assumption that mode_i_usage represents a specific mobility mode usage from all the available included modes examined (indicative array (public transport, taxi, car sharing, bike sharing)), while, similarly, Mode_i takes its values from the same array of available services.
Soft constraints are defined for all mobility modes involved in MaaS Plans and are specifically modified as per each city's provided modal allowances. Soft constraints are executed on the results delivered by the hard constraints and exclude MaaS Plans that contain mode levels that do not make sense for a specific user, e.g., when users indicate they make use of a particular mode of transport "Once/few times per month" (e.g., public transport), MaaS Plans that include maximum levels of that mode are excluded (e.g., for public transport, the maximum level stands for a 30-day public transport pass). Two indicative classes of soft constraints (SC 1 . . . n ) for monthly MaaS Plans are shown below: • SC 1 , "If user model.mode_i_usage = 'Once/few times per month', Mode_i ! = max values", conveying that, for users occasionally using a specific transport mode, MaaS Plans that have the maximum level of this particular transport mode are excluded. • SC 2 , "If user model.mode_i_usage = 'Once/few times per week', Mode_i ! = 0 and Mode_i ! = min values", denoting that users who are quite frequently using a specific transport mode, MaaS Plans that do not include that mode or have the minimum level of that mode are excluded.
We opted for maintaining two categories of constraints (hard and soft) for improved conceptualization, maintainability and testing. The formulation of the two constraint models is shown in Figure 3.
Similarity-Based Plan Ranking
The similarity-based Plan Ranking formula runs in the second step of the approach, processing the outcome of the CSP-based filtering. The goal is to order the MaaS Plans in regard to the user's profile, as this is shaped by the feedback s/he provides to the system. The formula considers the user's stated mobility habits with regard to the usage frequency of the transport modes that are part of the MaaS Plans as well as the user's stated willingness toward including the proposed modes within his/her MaaS Plan. Users' habits and willingness to include a specific mobility mode in a MaaS Plan are acquired through relevant questionnaires in the MaaS app (see Section 4.3). In more detail, two user vectors are considered: User-Habits and User-Willingness. A similarity value is then calculated between each one of the two aforementioned vectors and the filtered MaaS Plans.
The User-Habits vector represents users' transportation habits in the n-dimensional feature space, where n refers to the available transport modes. For instance, in a city were the modes considered are public transport (PT), taxi (TX), bike sharing (BS) and car sharing (CS), n equals to four as follows: In order to derive the values of the User-Habits vector, first, users are asked to state how often they use each mobility mode included within the examined MaaS Plans: (i) Never, (ii) Once/few times per month, (iii) Once/few times per week, (iv) Every day. In the MaaS Plans, there can be modes that commonly include an access pass with a specific duration (e.g., a daily or monthly pass for public transport or a daily or monthly pass for
Similarity-Based Plan Ranking
The similarity-based Plan Ranking formula runs in the second step of the approach, processing the outcome of the CSP-based filtering. The goal is to order the MaaS Plans in regard to the user's profile, as this is shaped by the feedback s/he provides to the system. The formula considers the user's stated mobility habits with regard to the usage frequency of the transport modes that are part of the MaaS Plans as well as the user's stated willingness toward including the proposed modes within his/her MaaS Plan. Users' habits and willingness to include a specific mobility mode in a MaaS Plan are acquired through relevant questionnaires in the MaaS app (see Section 4.3). In more detail, two user vectors are considered: User-Habits and User-Willingness. A similarity value is then calculated between each one of the two aforementioned vectors and the filtered MaaS Plans.
The User-Habits vector represents users' transportation habits in the n-dimensional feature space, where n refers to the available transport modes. For instance, in a city were the modes considered are public transport (PT), taxi (TX), bike sharing (BS) and car sharing (CS), n equals to four as follows: (1) In order to derive the values of the User-Habits vector, first, users are asked to state how often they use each mobility mode included within the examined MaaS Plans: (i) Never, (ii) Once/few times per month, (iii) Once/few times per week, (iv) Every day. In the MaaS Plans, there can be modes that commonly include an access pass with a specific duration (e.g., a daily or monthly pass for public transport or a daily or monthly pass for bike sharing) and modes that include distance-based quota (e.g., a predefined amount for taxi or car sharing use). For the latter modes, users are asked to provide an estimate of their average use in the form of average distance traveled per ride, e.g., average distance of taxi or car sharing rides. With the abovementioned information, the User-Habits vector is calculated based on the following formula: The frequencyFactor maps the frequency levels of the question asking users to state how often they use each mode to specific values, which can be defined with the support of transport engineers in order to represent the travel patterns of a specific area. The average_mode_usage is set to value one (1) for modes with access passes. For other modes, the user answer that indicates the average travel distance per ride with the specific mode is used. The max_mode_usage_frequency is a normalization factor calculated by multiplying the max frequencyFactor and max average_mode_usage.
By way of illustration, the frequencyFactor for a taxi service is set to {Never: 0; Once/few times per month: 3; Once/few times per week: 7; Every day: 22}, and the options for the average trip distance are set to 2 km, 3 km, 4 km, 5 km, 6 km. For a user that does not use taxi and selects "Never" when asked how often s/he uses taxi, the corresponding feature value of the User-Habits vector will be 0; for a user that uses taxi a few times per week for an average distance of 3km per ride, the corresponding feature value of the User-Habits vector will be 0.15.
In the same way, the User-Willingness vector is populated by the user's willingness to embody the various transport modes within a MaaS Plan. In particular, this vector is formed by the user's responses to the five-point Likert scale question, "Please define your willingness to include the following modes of transport in your new MaaS Plan", for the list of all the included transport services. The user's answers to the aforementioned question are normalized, and the User-Willingness vector is shaped as follows: Two cosine similarity coefficients are calculated between the two user vectors described above and the MaaS Plans vectors: In the final step, an aggregated similarity value is computed for each user and MaaS Plan, derived by the two aforementioned similarity coefficients. The list of ranked MaaS Plans constituting the result of the similarity mechanism is formed by sorting the MaaS Plans in descending order with respect to the aggregated similarity value. MaaS Plans presented at the top positions of the list are closer to the user's habits and willingness. The aggregated similarity formula is given below: similarityAggregated = similarity UserHabits−Plan 2 + similarity UserWillingness−Plan 2
Data-Driven Preferences Elicitation
Previous user selections regarding the subscribed MaaS Plans throughout his/her engagement with the MaaS app can be considered a valuable source of information that can be exploited and provide reasonable conclusions about the user's willingness to include the various mobility modes in a potential MaaS Plan they will choose within a next session. With this assumption in mind, a data-driven mechanism was designed to derive users' willingness to include specific transport modes. The mechanism is applied to already registered users in the MaaS app and, in particular, those who have already selected and used a MaaS subscription. For such users, the recommender system infers the user's willingness as the weighted average of the user-stated willingness and a system-deduced willingness from previous MaaS Plans subscriptions.
In more details, the system's inferred user willingness is calculated by considering the MaaS Plans the user has selected in the past, in particular, the transport modes and their related levels included within them. In the event of absence of a particular transport mode within the selected MaaS Plan, the system considers it as a sign of unwillingness to include that mode in future selections. On the contrary, the presence of a transport mode in maximum level is handled by the system as a high user willingness indication to also include it in a future session. The two aforementioned cases represent the two extremes, between which the rest of the intermediate mode levels of a transport mode are considered proportionally for the inclusion of that particular mode. A linear mapping between transport mode levels and willingness has been considered. It should be underlined that higher weights are allocated to latest plan subscription instances or user-stated willingness provisions, i.e., the approach contains a "forgetting factor" that assigns exponentially lower weights to older data (Sugiyama et al. [36]).
Consequently, the recommender constructs a willingness W vector considering a system-inferred willingness W system that indicates a specific user's willingness for all included modes, calculated by utilizing the user's past MaaS Plans subscriptions from n months ago and an explicitly stated user willingness W user . W user is essentially the vector composed by the user-stated willingness to include the available transport modes examined within the current MaaS schema, i.e., the vector −→ UW described in Section 3.2.
W system is established for all the available transport modes within the MaaS Plans by processing previous MaaS Plans subscriptions.
We use past subscriptions in an effort to build the W system and define S j (j = 0, 1, 2 . . . , n) as the set of past MaaS Plans subscriptions, which are accumulated within W system . Following this approach, each item is in the W system vector is specified as follows: where e − ln2×(d today −d i ) hl is a forgetting factor. More precisely, d i is the date when the subscription S i occurs, d today is the current date and hl is the half time parameter set to 30 days, denoting that diminishing factor by half in the period of one month. Additionally, we consider that a user has S n number of subscriptions. In conclusion, having formed W user and W system , the total willingness W is shaped as follows: where a and b are weighting factors that satisfy the equation a + b = 1 and allow to control the effect of each of the two vectors. In our approach, we set a = 0.7 and b = 0.3. An illustration of the aforementioned is depicted in Figure 4.
Controlled Experiment
The purpose of the controlled experiment was to examine and compare the performance of the proposed recommendation approach against a range of other approaches and understand how these affect users' decisions in the MaaS Plan selection process. For the purposes of the controlled experiment, we set up a web application that consolidated a set of MaaS Plans recommendation approaches, simulating the process of the plans selection task. The web application provides an adaptable framework within which the testing, comparison and evaluation of all the included algorithms are done, without the necessity for a real-life MaaS application.
Experimental Conditions-Recommendation Approaches
For the conduction of the experiment, a range of recommendation approaches that provide lists of MaaS Plans were chosen to be compared and evaluated. The intention was to infer conclusions regarding which algorithm provides MaaS Plans and corresponding lists that adhere to end-user needs and preferences.
The included recommendation approaches are described below: 1. Price-desc/asc, signifies two approaches that use the basic technique of price ranking of the available MaaS Plans either in descending or ascending order. 2. CSP approach, denotes the filtering approach described in Section 3.1, CSP which contains hard and soft constraints that are independently implemented with the latter performing on top of the results of the first. Finally, on the subset of filtered plans delivered by the CSP, a price ranking in ascending order is used. 3. CSP with similarity approach, in this case, filtering based on hard and soft constraints is performed, whereas as a final step the similarity-based Plan Ranking technique is used on top of the filtered products and delivers a ranked list of MaaS Plans where, in the top positions, those most similar to the user profile plans are presented. 4. CSP with similarity and price filter, the final approach includes approach 3 described above, along with an extra feature of price filtering, allowing users to adjust the plans within a restricted budget. Price filtering is a popular feature among e-commerce ap-
Controlled Experiment
The purpose of the controlled experiment was to examine and compare the performance of the proposed recommendation approach against a range of other approaches and understand how these affect users' decisions in the MaaS Plan selection process. For the purposes of the controlled experiment, we set up a web application that consolidated a set of MaaS Plans recommendation approaches, simulating the process of the plans selection task. The web application provides an adaptable framework within which the testing, comparison and evaluation of all the included algorithms are done, without the necessity for a real-life MaaS application.
Experimental Conditions-Recommendation Approaches
For the conduction of the experiment, a range of recommendation approaches that provide lists of MaaS Plans were chosen to be compared and evaluated. The intention was to infer conclusions regarding which algorithm provides MaaS Plans and corresponding lists that adhere to end-user needs and preferences.
The included recommendation approaches are described below: 1.
Price-desc/asc, signifies two approaches that use the basic technique of price ranking of the available MaaS Plans either in descending or ascending order.
2.
CSP approach, denotes the filtering approach described in Section 3.1, CSP which contains hard and soft constraints that are independently implemented with the latter performing on top of the results of the first. Finally, on the subset of filtered plans delivered by the CSP, a price ranking in ascending order is used.
3.
CSP with similarity approach, in this case, filtering based on hard and soft constraints is performed, whereas as a final step the similarity-based Plan Ranking technique is used on top of the filtered products and delivers a ranked list of MaaS Plans where, in the top positions, those most similar to the user profile plans are presented. 4.
CSP with similarity and price filter, the final approach includes approach 3 described above, along with an extra feature of price filtering, allowing users to adjust the plans within a restricted budget. Price filtering is a popular feature among e-commerce applications, which may likely be advantageous in the MaaS Plans selection problem. The feature concerns a specific functionality that provides users the ability to adjust the proposed product assortments within a budget they define and which practically filters MaaS Plans within a user's price constraints. The user has the option to tune the price filter according to his/her budget in order to filter out plans that are priced higher. Table 1 summarizes the different approaches described above, along with their corresponding features. In terms of the experimental settings, we followed a within-subjects design commonly used in recommender systems evaluations (see, e.g., Ekstrand et al. [37], Paolacci et al. [38]), where participants are presented with pairs of recommendation lists generated by different approaches. The within-subjects approach allowed us to gather more results from the participants and minimize random noise, as explained in Charness et al. [39]. Moreover evaluation is a certainly comparable task, and evaluating each algorithm individually would deprive the users from relating them to one another as stated in Hsee, & Zhang [40]. Note that the pair "Price-asc" versus "Price-desc" was excluded from the list of potential joint evaluations, as there was no point to comparing these two extremes.
Users and Context
The experiment was deployed and instantiated in three different cities. The first instance was configured for the city of Budapest in Hungary, where MaaS pilots are being deployed and an introduction of a related concept has been already initiated to city inhabitants. Participants included graduate and postgraduate students from the Budapest University of Technology and Economics. A total of 302 users were recruited to participate, out of which 110 successfully completed the survey. The participants' ages ranged between 21 and 39 years old. The second instance was configured for the area of California in the US and its corresponding mobility options. Users were recruited through the popular crowdsourcing platform Amazon Mechanical Turk (MTurk), which, as stated by Paolacci et al. [38], is viewed as a very practical option for data collection, considering that the participants demonstrate typical heuristics and biases and concentrate to guidelines no less than traditional sources' participants do. Location-based restrictions were applied in MTurk to ensure that only users from the area of California could participate. In total, 268 users took part, out of which 113 completed the whole cycle of the experiment. In this instance, the ages of the participants ranged between 18 and 69 years old. The third instance was configured for the area of Greater Manchester in the UK. The participants were recruited by Atkins, a company that provides surveying services, and the participants' ages were between 18 and 74 years old. There were 88 attempts, from which 39 were completed successfully.
Experiment Survey and Process
The experiment's survey was developed as a web application, and users had to complete a five-step process. The survey begins with an introductory text and an informative video ( Figure 5) that explains the concept of MaaS and tries to familiarize users with MaaS terms, including basic instructions of how to proceed with the experiment. Users press the "Start" button and move to the second step, where a group of sociodemographic questions are presented, intending to construct a user profile for the purposes of the current experimental session. The questions ask participants to provide personal information about their gender, their level of education, their working schedule, their employment status and to indicate if they already own a car and a driving license. The set of questions is displayed in Figure 6a. Subsequently, in the third step of the process, users are requested to complete a set of questions about their mobility habits. The corresponding questions ask users about the frequency of usage of each transport mode included in the MaaS instance examined, indicatively, "How often do you use public transport?" The user answers are used to construct the User-Habits vector; thus, the present questions are mandatory to answer. An indicative example of the set of questions is illustrated in Figure 6b. The various modes of transport included in this set of questions diversify according to the area where the experiment is deployed and conform with the mobility services available in this specific area. In the next step and within the "User Mode Preferences" page, the user is asked to state her/his willingness to include the available modes of transport in their MaaS Plan ("Please define your willingness to include the following modes of transport in your new MaaS Plan"). The potential answers are all shaped in a 1-5 Likert scale as it is displayed in Figure 6c per available transport mode. Figure 6b. The various modes of transport included in this set of questions diversify according to the area where the experiment is deployed and conform with the mobility services available in this specific area. In the next step and within the "User Mode Preferences" page, the user is asked to state her/his willingness to include the available modes of transport in their MaaS Plan ("Please define your willingness to include the following modes of transport in your new MaaS Plan"). The potential answers are all shaped in a 1-5 Likert scale as it is displayed in Figure 6c per available transport mode. Upon completion of all the questions described above, the system has collected all the essential data for the recommender service to run and produce the ranked list of MaaS Plans fitting a particular user's needs. Prior to displaying the ranked lists of MaaS Plans to the users, a short explanatory message is displayed (Figure 7), clarifying to the user what s/he will be presented with and how s/he is asked to evaluate the given outcome of the system. Moreover, a table with the various levels of the included transport modes is presented to the participant, listing the modal allowances the MaaS Plans consist of. As exemplified in Figure 7, the user is informed about the diverse classes (xsmall, small, medium, large, xlarge) that describe the transport modes, which will be provided according to the user's level of usage. Upon completion of all the questions described above, the system has collected all the essential data for the recommender service to run and produce the ranked list of MaaS Plans fitting a particular user's needs. Prior to displaying the ranked lists of MaaS Plans to the users, a short explanatory message is displayed (Figure 7), clarifying to the user what s/he will be presented with and how s/he is asked to evaluate the given outcome of the system. Moreover, a table with the various levels of the included transport modes is presented to the participant, listing the modal allowances the MaaS Plans consist of. As exemplified in Figure 7, the user is informed about the diverse classes (xsmall, small, medium, large, xlarge) that describe the transport modes, which will be provided according to the user's level of usage. After the user finishes reading the aforementioned text, two lists of recommended MaaS Plans appear following the within-subjects evaluation. The pair of approaches is randomly selected each time in a way that ascertains that each approach is presented an equal number of times. Each of the presented lists shows five different MaaS Plans within a table, the rows of which correspond to a different MaaS Plan, and the columns depict the transport mode and the suggested mode level value, along with the price range of the proposed plan. Moreover, a rating column with a 5-star scale is shown, asking the user to rate the presented plans based on their preference. Figure 8 provides an indicative view of the plans recommendation lists for the recommendation approach "CSP with similarity and price filter option" (List 1) and the approach "Price-ascending" (List 2). After having rated the recommended MaaS Plans of each one of the two lists, an "Exit Questionnaire" follows, where the user is asked to respond to a set of questions assessing different qualities of the recommendations and the adoption of MaaS. In more detail, the questions are displayed in Figure 9. The participants' responses to all the above-mentioned questions, along with the plans' ratings, are further explored in the course of the evaluation stage of the experiment. After the user finishes reading the aforementioned text, two lists of recommended MaaS Plans appear following the within-subjects evaluation. The pair of approaches is randomly selected each time in a way that ascertains that each approach is presented an equal number of times. Each of the presented lists shows five different MaaS Plans within a table, the rows of which correspond to a different MaaS Plan, and the columns depict the transport mode and the suggested mode level value, along with the price range of the proposed plan. Moreover, a rating column with a 5-star scale is shown, asking the user to rate the presented plans based on their preference. Figure 8 provides an indicative view of the plans recommendation lists for the recommendation approach "CSP with similarity and price filter option" (List 1) and the approach "Price-ascending" (List 2). After having rated the recommended MaaS Plans of each one of the two lists, an "Exit Questionnaire" follows, where the user is asked to respond to a set of questions assessing different qualities of the recommendations and the adoption of MaaS. In more detail, the questions are displayed in Figure 9. The participants' responses to all the above-mentioned questions, along with the plans' ratings, are further explored in the course of the evaluation stage of the experiment. Finally, by pressing the "Save My Preferences" button, the experiment session ends, and the system saves users' input, including his/her profile, the two distinct algorithms that are displayed each time within the running session, the start and end datetime of the conduction of the experiment, the two lists along with the ranked MaaS Plans displayed per each algorithm, the ratings of the MaaS Plans provided by the user and the users' answers to the qualitative questionnaire.
Results
The results of the controlled experiment were extracted from two sources: (i) users' answers to the question about their choice between the two presented lists, along with the average rating of the MaaS Plans on a list basis (i.e., the preferred recommendation approach), and (ii) the user answers to the five-point Likert scale questions of the exit questionnaire. Before the analysis, data were cleaned in order to remove invalid answers. More specifically, the average time execution of the experiment was calculated in 4 to 4.5 min, shaping the experiment's median completion time. We removed cases of participants that conducted the survey in less than 1.5 or more than 12 min, as participants provided in- Finally, by pressing the "Save My Preferences" button, the experiment session ends, and the system saves users' input, including his/her profile, the two distinct algorithms that are displayed each time within the running session, the start and end datetime of the conduction of the experiment, the two lists along with the ranked MaaS Plans displayed per each algorithm, the ratings of the MaaS Plans provided by the user and the users' answers to the qualitative questionnaire.
Results
The results of the controlled experiment were extracted from two sources: (i) users' answers to the question about their choice between the two presented lists, along with the average rating of the MaaS Plans on a list basis (i.e., the preferred recommendation approach), and (ii) the user answers to the five-point Likert scale questions of the exit questionnaire. Before the analysis, data were cleaned in order to remove invalid answers. More specifically, the average time execution of the experiment was calculated in 4 to 4.5 min, shaping the experiment's median completion time. We removed cases of participants that conducted the survey in less than 1.5 or more than 12 min, as participants provided incomplete responses in these cases. Moreover, we checked the remaining cases and removed those where mismatches were observed following the answer to control question. In particular, we identified and removed cases where users provided high ratings to the plans of one of the two recommendation lists and subsequently stated that they prefer the other list in the control question. In the final dataset, the CSP approach was presented 101 times, the CSP_Sim was displayed 138 times, the CSP-Filter was shown 110 times, the Price-asc approach 94 times and the Price-desc approach 81 times. Figure 10 depicts users' answers to the question, "Please select the list that you would choose in a real-life MaaS setting", in terms of percentages for the different approaches. The "CSP with similarity" approach was selected in 75% of the cases where it was presented either in List 1 or List 2. The approach, "CSP with similarity and price filter option", is placed second and was preferred in 50% of the cases. This finding amplifies the role of CSP filtering, since the two most preferred approaches embed this filtering method.
complete responses in these cases. Moreover, we checked the remaining cases and removed those where mismatches were observed following the answer to control question. In particular, we identified and removed cases where users provided high ratings to the plans of one of the two recommendation lists and subsequently stated that they prefer the other list in the control question. In the final dataset, the CSP approach was presented 101 times, the CSP_Sim was displayed 138 times, the CSP-Filter was shown 110 times, the Price-asc approach 94 times and the Price-desc approach 81 times. Figure 10 depicts users' answers to the question, "Please select the list that you would choose in a real-life MaaS setting", in terms of percentages for the different approaches. The "CSP with similarity" approach was selected in 75% of the cases where it was presented either in List 1 or List 2. The approach, "CSP with similarity and price filter option", is placed second and was preferred in 50% of the cases. This finding amplifies the role of CSP filtering, since the two most preferred approaches embed this filtering method. A more thorough analysis on how the proposed "CSP with similarity formula" performed over the remaining approaches is provided in Table 2. The first observation conveys that, in the case of "CSP_Sim" versus "Price-descending" pair, the CSP_Sim was selected 80% of the time. Next, for the case of the "CSP_Sim" versus "Price-asc" pair, the "CSP_Sim" was also preferred over 82.86% of the time, whereas, for the case "CSP_Sim" presented with "CSP", the "CSP with similarity" was chosen over 57.14% of the time. Conclusively, for the case of "CSP with similarity" versus "CSP with similarity and price filter" algorithm, the CSP_Sim was preferred in over 80% of the cases. Table 2. CSP_Sim selected when presented with the remaining algorithms.
Presented Pairs
CSP_Sim Selected (%) CSP with similarity vs. Price-descending 80% CSP with similarity vs. Price-ascending 82.86% CSP with similarity vs. CSP 57.14% CSP with similarity vs. CSP with similarity and price filter 80% Table 3 provides a summarized view of the average star ratings of the MaaS Plans included in the CSP_Sim assortments against those derived from the other algorithms, as they are presented in pairs each time. The MaaS Plans calculated with the "CSP_Sim" approach were, in all cases, rated higher. A more thorough analysis on how the proposed "CSP with similarity formula" performed over the remaining approaches is provided in Table 2. The first observation conveys that, in the case of "CSP_Sim" versus "Price-descending" pair, the CSP_Sim was selected 80% of the time. Next, for the case of the "CSP_Sim" versus "Price-asc" pair, the "CSP_Sim" was also preferred over 82.86% of the time, whereas, for the case "CSP_Sim" presented with "CSP", the "CSP with similarity" was chosen over 57.14% of the time. Conclusively, for the case of "CSP with similarity" versus "CSP with similarity and price filter" algorithm, the CSP_Sim was preferred in over 80% of the cases. Table 2. CSP_Sim selected when presented with the remaining algorithms.
Presented Pairs CSP_Sim Selected (%)
CSP with similarity vs. Price-descending 80% CSP with similarity vs. Price-ascending 82.86% CSP with similarity vs. CSP 57.14% CSP with similarity vs. CSP with similarity and price filter 80% Table 3 provides a summarized view of the average star ratings of the MaaS Plans included in the CSP_Sim assortments against those derived from the other algorithms, as they are presented in pairs each time. The MaaS Plans calculated with the "CSP_Sim" approach were, in all cases, rated higher. Note that the average star rating per list was formed as the average rating of the MaaS Plans contained within the particular list. T-test analysis was performed among the ratings of the different pairs in order to estimate the statistical significance between the mean ratings. The results showed statistical significance of the following pairs: "CSP with similarity vs. Price-descending" (t = −2.69, p < 0.05), "CSP with similarity vs. Priceascending" (t = −3.54, p < 0.05) and "CSP with similarity vs. CSP" (t = −2.02, p < 0.1). The difference in the case of "CSP with similarity vs. CSP with similarity and price filter" was not found statistically significant for the given sample sizes.
Regarding the users' responses to the five-point Likert scale exit questionnaire that was presented and filled in at the end of the evaluation process, the results are shown in Figure 11. The specific questions asked users to choose between two lists and identify how their choice better aligned with their preferences. More precisely, users were requested to provide their feedback on a five-point Likert scale, annotated on the left side with "Much more List A", and on the right side with "Much more List B" (Figure 9). The corresponding results per question are illustrated in histograms containing normalized values and grouped user responses as follows: List A is considered to better comply with the presented statement when the users' answers are either 1 or 2, whereas List B is considered preferred when the response is either 4 or 5. Note that the average star rating per list was formed as the average rating of the MaaS Plans contained within the particular list. T-test analysis was performed among the ratings of the different pairs in order to estimate the statistical significance between the mean ratings. The results showed statistical significance of the following pairs: "CSP with similarity vs. Price-descending" (t = −2.69, p < 0.05), "CSP with similarity vs. Price-ascending" (t = −3.54, p < 0.05) and "CSP with similarity vs. CSP" (t = −2.02, p < 0.1). The difference in the case of "CSP with similarity vs. CSP with similarity and price filter" was not found statistically significant for the given sample sizes.
Regarding the users' responses to the five-point Likert scale exit questionnaire that was presented and filled in at the end of the evaluation process, the results are shown in Figure 11. The specific questions asked users to choose between two lists and identify how their choice better aligned with their preferences. More precisely, users were requested to provide their feedback on a five-point Likert scale, annotated on the left side with "Much more List A", and on the right side with "Much more List B" (Figure 9). The corresponding results per question are illustrated in histograms containing normalized values and grouped user responses as follows: List A is considered to better comply with the presented statement when the users' answers are either 1 or 2, whereas List B is considered preferred when the response is either 4 or 5. Concerning the question, "Which list has more MaaS Plans that are close to your preferences?", the "CSP with similarity" approach is in first place, while the "CSP" is in second place. More specifically, lists generated with "CSP_Sim" were selected 63% of the time that they were displayed, while lists generated with the "CSP" approach were selected 53% of the time. In the question, "Which list has more plans that you find appealing?", "CSP with similarity" is placed first with 57%, while the "CSP-Filter" algorithm is second (50%) and in the third place, the "CSP" together with "Price-asc" is observed with 48%. Finally, the user responses to the question, "Which list has more plans that are in line with your budget for transportation?", shows that lists generated with "Price-asc" were the most desired (57%); thereafter, the lists generated with "CSP" approach follow with 56%, while lists with "CSP with price filter" come in the third place (50%). An interpretation of these results is that, under for this particular question, users consider certain budget constraints, which reasonably leads them to less expensive plans, which can be found most easily within Price-asc lists or in the CSP algorithm, where the plans were also Which list has more plans that are inline with your budget for transportation? Figure 11. The results of the Exit Questionnaire.
Concerning the question, "Which list has more MaaS Plans that are close to your preferences?", the "CSP with similarity" approach is in first place, while the "CSP" is in second place. More specifically, lists generated with "CSP_Sim" were selected 63% of the time that they were displayed, while lists generated with the "CSP" approach were selected 53% of the time. In the question, "Which list has more plans that you find appealing?", "CSP with similarity" is placed first with 57%, while the "CSP-Filter" algorithm is second (50%) and in the third place, the "CSP" together with "Price-asc" is observed with 48%. Finally, the user responses to the question, "Which list has more plans that are in line with your budget for transportation?", shows that lists generated with "Price-asc" were the most desired (57%); thereafter, the lists generated with "CSP" approach follow with 56%, while lists with "CSP with price filter" come in the third place (50%). An interpretation of these results is that, under for this particular question, users consider certain budget constraints, which reasonably leads them to less expensive plans, which can be found most easily within Price-asc lists or in the CSP algorithm, where the plans were also ranked in ascending price order. Moreover, the price filter option gained significant attention, since it provided the opportunity to adjust the suggested MaaS Plans assortments within the user's subjective price budget. Figure 12 illustrates the answers to the question, "What were the three main reasons for rating higher MaaS Plans". The majority of the participants rated higher what resembled "Most similar to what I use today", then "Best transport mode combinations" is in second place with 26%. Next, the reason "Cheapest" follows, with 24%, and finally, whether a particular MaaS Plan was of the "Best value" is considered as the last reason that affected the users choice to rate this plan higher. The answers to this question highlight the fact that properly identifying users' habits and transport mode preferences is crucial in the process of recommending meaningful and preferable MaaS Plans. ranked in ascending price order. Moreover, the price filter option gained significant attention, since it provided the opportunity to adjust the suggested MaaS Plans assortments within the user's subjective price budget. Figure 12 illustrates the answers to the question, "What were the three main reasons for rating higher MaaS Plans". The majority of the participants rated higher what resembled "Most similar to what I use today", then "Best transport mode combinations" is in second place with 26%. Next, the reason "Cheapest" follows, with 24%, and finally, whether a particular MaaS Plan was of the "Best value" is considered as the last reason that affected the users choice to rate this plan higher. The answers to this question highlight the fact that properly identifying users' habits and transport mode preferences is crucial in the process of recommending meaningful and preferable MaaS Plans. Moreover, we examined the stated reasons for rating higher MaaS Plans while considering the sociodemographic characteristics of the respondents, aiming to identify differences across user groups. The results showed that full-time-employed participants considered the "Best value" of MaaS Plans in contrast to participants with those with employment statuses of part-time employed, retired or student, whose choices were determined by "Best transport mode combinations". The differences could be attributed to the fact that full-time-employed persons commute daily and seek to balance cost and benefits. Differences were also observed when we considered the age of the participants. In particular, participants in the age group of 54-72 were more inclined to rate higher MaaS Plans "Most similar to what I use today". This could be attributed to the fact that such persons are not willing to change their current habits. No major differences were observed when considering the sex and education of the participants.
Additional ideas stated by the participants of the controlled experiment were their replies to the query, "What other information would you like to see in order to support you better to choose MaaS Plans?" The findings are provided in Table 4, below. Moreover, we examined the stated reasons for rating higher MaaS Plans while considering the sociodemographic characteristics of the respondents, aiming to identify differences across user groups. The results showed that full-time-employed participants considered the "Best value" of MaaS Plans in contrast to participants with those with employment statuses of part-time employed, retired or student, whose choices were determined by "Best transport mode combinations". The differences could be attributed to the fact that full-timeemployed persons commute daily and seek to balance cost and benefits. Differences were also observed when we considered the age of the participants. In particular, participants in the age group of 54-72 were more inclined to rate higher MaaS Plans "Most similar to what I use today". This could be attributed to the fact that such persons are not willing to change their current habits. No major differences were observed when considering the sex and education of the participants.
Additional ideas stated by the participants of the controlled experiment were their replies to the query, "What other information would you like to see in order to support you better to choose MaaS Plans?" The findings are provided in Table 4, below. The results show that users would like to have access to additional information about the terms and conditions of the transport modes included within the purchased MaaS Plans, e.g., wi-fi available, etc., or certain instructions of how the mode allowances of the modes should be used (e.g., 1 h car sharing). Moreover, the users would be interested to know the individual prices of the modes if these were not included in a MaaS Plan. There is also a keen interest by the users for more transport services to be included (e.g., parking, e-scooter) or to enjoy the experience of creating their own MaaS Plan. Other propositions include MaaS Plans without public transport services and the promotion of MaaS through potential awards/bonuses.
Real Life Pilots
The proposed plans recommendation approach was integrated in a mobile application developed to support real life MaaS pilots in the cities of Budapest, Manchester and Luxemburg as part of a collaborative European research project called MaaS4EU. The MaaS4EU mobile app was available in Google Play Store and App Store for both Android and iOS devices, providing a range of options to the user that facilitate the adoption of MaaS. The MaaS Plans selection process within the app is illustrated in Figure 13. The process starts with a set of questionnaires that extract the user's mobility habits, asking about the usage frequency of the various transport modes included in the examined MaaS schema (e.g., "How often do you use public transport?") and other information that shapes a user's mobility profile (e.g., "Do you have a full driving license?", "Does your household own one or more cars?"). In the next step, users are asked to indicate their willingness to include the specific available transport modes within their MaaS Plan. At that point, all the required user input has been collected and processed. The recommender runs, incorporating the user input within the predefined constraints and the similarity formula. The system provides a ranked list of suggested MaaS Plans. The top three recommended plans are then displayed, while the user is provided with the option of "Load More" to further browse more plans. The user selects the MaaS Plan that attracts her/him more and has access to a more detailed description, indicating the various levels of the transport modes within each specific MaaS Plan. Finally, the user may decide to buy a specific plan and subscribe to MaaS. Note that, for the case of an existing user that has already provided her/his feedback into the system, s/he is asked during the initial launch of the app whether s/he would prefer to update her/his preferences. Alternatively, the system executes on top of the already registered user's data input. Concerning the performance of the proposed recommender service that was integrated and utilized within the mobile app version for the pilot phase, an interesting query that would estimate its efficiency within each MaaS setting and city was the position of the selected MaaS Plan within the presented ranked list of plans. The goal is to present the actual purchased MaaS Plans by the users within the top positions of the list, meaning that the system has the ability to understand a user's special needs and thus provide the desired MaaS product presented in a distinct place. Figure 14 indicates that the majority of the selected/purchased MaaS Plans were successfully chosen from the top three places in all cases. Note that, in the case of Budapest, the pilot was split in two periods. The first one started in January 2020 but was abruptly stopped due to the COVID-19 pandemic. The pilot restarted in August 2020. Concerning the performance of the proposed recommender service that was integrated and utilized within the mobile app version for the pilot phase, an interesting query that would estimate its efficiency within each MaaS setting and city was the position of the selected MaaS Plan within the presented ranked list of plans. The goal is to present the actual purchased MaaS Plans by the users within the top positions of the list, meaning that the system has the ability to understand a user's special needs and thus provide the desired MaaS product presented in a distinct place. Figure 14 indicates that the majority of the selected/purchased MaaS Plans were successfully chosen from the top three places in all cases. Note that, in the case of Budapest, the pilot was split in two periods. The first one started in January 2020 but was abruptly stopped due to the COVID-19 pandemic. The pilot restarted in August 2020.
The results show that, in the first period of the pilot in Budapest, 79% of the purchased Plans were included in the top three positions, while for the second period, the purchased MaaS Plans were in the top three positions in 82% of the cases. In the case of Greater Manchester, the selected plans were in the top three positions in 81% of the cases, whereas in the city of Luxemburg, the proportion was 64%.
Last but not least, a supplementary mechanism that intended to gain explicit user feedback about the recommended MaaS Plans was implemented in the form of notification messages that were displayed to the users of the app when they navigated in the plan selection screen (Figure 15). Two messages were defined and integrated in the notifications mechanism, asking users to provide their degree of agreement or disagreement in a five-point Likert scale extending from Strongly disagree (1) to Strongly Agree (5), to the following questions: "The recommended plans are close to my travel preferences" (Q1) and "The recommended plans include transport modes that match my preferences" (Q2). Sustainability 2021, 13, x FOR PEER REVIEW 25 of 29 The results show that, in the first period of the pilot in Budapest, 79% of the purchased Plans were included in the top three positions, while for the second period, the purchased MaaS Plans were in the top three positions in 82% of the cases. In the case of Greater Manchester, the selected plans were in the top three positions in 81% of the cases, whereas in the city of Luxemburg, the proportion was 64%.
Last but not least, a supplementary mechanism that intended to gain explicit user feedback about the recommended MaaS Plans was implemented in the form of notification messages that were displayed to the users of the app when they navigated in the plan selection screen (Figure 15). Two messages were defined and integrated in the notifications mechanism, asking users to provide their degree of agreement or disagreement in a five-point Likert scale extending from Strongly disagree (1) to Strongly Agree (5), to the following questions: "The recommended plans are close to my travel preferences" (Q1) and "The recommended plans include transport modes that match my preferences" (Q2). Figure 16 shows the results of the users' answers. The results indicate that most users agreed or strongly agreed that the recommended plans were close to their preferences and included transport modes that they prefer. Figure 16 shows the results of the users' answers. The results indicate that most users agreed or strongly agreed that the recommended plans were close to their preferences and included transport modes that they prefer. Figure 15. Indicative notification message intended to gain explicit user feedback about the recommended MaaS Plans. Figure 16 shows the results of the users' answers. The results indicate that most users agreed or strongly agreed that the recommended plans were close to their preferences and included transport modes that they prefer.
Conclusions
In this paper, we presented the MaaS Plans Recommender, designed and implemented in order to support MaaS end-users in identifying and selecting the mobility plans that fit their transportation needs. The proposed recommender provides filtering functionalities that rely on concepts of constraint programming by leveraging user feedback
Conclusions
In this paper, we presented the MaaS Plans Recommender, designed and implemented in order to support MaaS end-users in identifying and selecting the mobility plans that fit their transportation needs. The proposed recommender provides filtering functionalities that rely on concepts of constraint programming by leveraging user feedback in a knowledge-based implementation. This approach was chosen due to its ability to tackle the so-called cold start problem, which is apparent within new fields of research or market that lack past data, including MaaS. Moreover, the recommender ranks filtered MaaS Plans with the use of a similarity formula, which considers users' habits with respect to the use of different transport modes as well as their willingness to include different transport modes in their plan. When past user choices are available, the recommender considers them and infers users' preferences in a data-driven manner. The proposed recommender was evaluated in experimental settings as well as in real life situations in the context of MaaS pilots, which were deployed in Budapest (Hungary), Luxemburg and Greater Manchester (UK). The experimental results showed that the proposed approach provides lists of MaaS Plans that users would choose in a real-life MaaS setting, in the majority of the cases. Moreover, the results of the real-life pilots showed that most of the participants chose an actual MaaS Plan from the top three places of the recommendation lists.
The proposed recommender can be utilized and fit into potential MaaS applications. When deploying the recommender, practitioners need to perform proper configuration of the recommendation service by performing a thorough study of available transport modes and MaaS Plans. In particular, the available mobility modes and their attributes are the elements that form the constraints of the recommendation model and should be configured and updated for the application at hand. Moreover, the mechanism that infers the similarity between user preferences and MaaS Plans and relies on capturing users' habits and willingness to include different modes of transport in a MaaS Plan. As the ranges of the various mobility modes included in MaaS Plans depends on the application at hand, the corresponding hard coded information needs to be modeled and configured into the RS on a per city case basis.
Last but not least, future research is needed to examine the following aspects of the proposed approach. First, as already described, the proposed recommender system integrates a data-driven module that infers users' preferences based on past plans choices. Our pilot studies could not capture enough repeated user choices that would allow us to gather enough data to properly evaluate this aspect of our approach. Longer term and longitudinal studies are needed so that users purchase enough subscriptions and the available data are adequate for generating recommendations based on past user choices, which would allow for proper evaluation of this aspect. Moreover, such studies could be used to understand the effects of seasonality and how user choices change in different times of the year, thus informing the recommendation process. Furthermore, in our approach, we considered item-based similarity measures to infer the similarity between user preferences and MaaS Plans. As MaaS becomes mainstream and more data are available, future research should focus on examining user-based similarity measures that could uncover mobility habits that similar users present, shaping potential clusters of communities within the MaaS schema. Additionally, metrics such as distance, cost, safety and traffic could potentially be incorporated within future versions of the recommender system and amplify its personalization ability. | 17,896.2 | 2021-07-23T00:00:00.000 | [
"Computer Science",
"Engineering",
"Environmental Science"
] |
Statistical Coalescence Model Analysis of J/psi Production in Pb+Pb Collisions at 158 A GeV
Production of J/\psi mesons in heavy ion collisions is considered within the statistical coalescence model. The model is in agreement with the experimental data of the NA50 Collaboration for Pb+Pb collisions at 158 A GeV in a wide centrality range, including the so called ``anomalous'' suppression domain. The model description of the J/psi data requires, however, strong enhancement of the open charm production in central Pb+Pb collisions. This model prediction may be checked in the future SPS runs.
Production of charmonium states J/ψ and ψ ′ in nucleus-nucleus collisions has been studied at CERN SPS over the previous 15 years by the NA38 and NA50 Collaborations. This experimental program was mainly motivated by the suggestion [1] to use the J/ψ as a probe of the state of matter created at the early stage of the collision. The original picture [1] (see also [2] for a modern review) assumes that charmonia are created exclusively at the initial stage of the reaction in primary nucleon-nucleon collisions. During the subsequent evolution of the system, the number of hidden charm mesons is reduced because of: (a) absorption of pre-resonance charmonium states by nuclear nucleons (normal nuclear suppression), (b) interactions of charmonia with secondary hadrons (comovers), (c) dissociation of cc bound states in deconfined medium (anomalous suppression). It was found [3] that J/ψ suppression with respect to Drell-Yan muon pairs measured in proton-nucleus and nucleus-nucleus collisions with light projectiles can be explained by the so called "normal" (due to sweeping nucleons) nuclear suppression alone. In contrast, the NA50 experiment with a heavy projectile and target (Pb+Pb) revealed essentially stronger J/ψ suppression for central collisions [4][5][6][7]. This anomalous J/ψ suppression was attributed to formation of quark-gluon plasma (QGP) [7], but a comover scenario cannot be excluded [8].
A completely different picture of charmonium production was developed recently within several model approaches [9][10][11][12][13][14]. In contrast to the standard approach, hidden charm mesons are supposed to be created at the hadronization stage of the reaction due to coalescence of c andc quarks created earlier. In this case the J/ψ yield is not restricted from above by the normal nuclear suppression curve. Therefore, neither anomalous suppression nor enhancement are excluded.
In the present letter we consider the statistical coalescence model (SCM) [10,11] of charmonium production. We assume that c andc are created at the initial stage of the reaction in primary hard parton collisions. We neglect creation of cc pairs after the hard initial stage as well as their possible annihilation. Then, the number of charmed quark-antiquark pairs remain approximately unchanged during subsequent stages. They are distributed over final hadron states at the hadronization stage in accord with laws of statistical mechanics. The SCM provides an excellent quantitative description of the NA50 data on centrality dependence of J/ψ production in Pb+Pb collisions at SPS, provided that the number of nucleon participants is not too small (N p > ∼ 100). The peripheral collision data can be explained qualitatively.
If creation of heavy quarks is indeed a hard process only, the average number cc AB(b) of produced cc pairs must be proportional to the number of primary nucleon-nucleon collisions. Then the centrality dependence of cc AB(b) can be calculated in Glauber's approach: Here b is the impact parameter, T AB (b) is the nuclear overlap function (see Appendix) and σ N N cc is the cc production cross section for nucleon-nucleon collisions. As discussed in Ref. [15], deconfined medium can substantially modify charm production in hard collisions at SPS. Therefore, σ N N cc in A+B collisions can be different from the corresponding cross section measured in a nucleon-nucleon collision experiment. The present analysis considers σ N N cc as a free parameter. Its value is fixed by fitting the NA50 data. Event-by-event fluctuations of the number of cc pairs follow the binomial distribution, which can be safely approximated by the Poisson distribution because the probability to produce a cc pair in a nucleon-nucleon collision is small: where P k (b) is the probability to produce k cc pairs in an A+B collision at impact parameter b. Assuming exact cc-number conservation during the evolution of the system, the SCM result for the average number of produced J/ψ per A+B collision is given by [11] Here is the total open charm thermal multiplicity. The sum runs over all known (anti)charmed particle species [16]. The total J/ψ multiplicity includes the contribution of excited charmonium states decaying into J/ψ: Here R(j) is the decay branching ratio of the charmonium j into J/ψ: R(J/ψ) ≡ 1, R(χ 1 ) ≈ 0.27, R(χ 2 ) ≈ 0.14 and R(ψ ′ ) ≈ 0.54. The multiplicities N j are found in the grand canonical ensemble formulation of the equilibrium hadron gas model: Here V and T are the volume 1 and temperature of the HG system, m j and d j denote, respectively, the masses and degeneracy factors of particles. K 2 is the modified Bessel function. The chemical potential µ j of the particle species j in Eq.(6) is defined as Here b j ,s j and c j represent the baryon number, strangeness and charm of the particle j, respectively. The baryonic chemical potential µ B regulates the baryonic density. The strange µ S and charm µ C chemical potentials are found by requiring zero value for the total strangeness and charm in the system . In our consideration we neglect small effects of a non-zero electrical chemical potential. We assume that the chemical freeze-out occurs close to (or even coincide with) the hadronization (where charmonia supposedly formed). Therefore for the thermodynamic parameters T and µ B we use the chemical freeze-out values found [17] by fitting the HG model to the hadron yield data in Pb+Pb collisions at SPS: Uncertainties in the freeze-out parameters exist due to time evolution of the system through the phase transition [18] and because of possible change of effective hadron masses in hot and dense hadron medium [19]. To check robustness of the predictions, an independent parameter set [20] is also used. It has been obtained by assuming strangeness and antistrangeness suppression by factor γ s : The system is assumed to freeze-out chemically at some common volume. This is fixed by the condition of baryon number conservation: Here Eq. (3) gives the total number of produced J/ψ-s. They decay into µ + µ − with the probability B J/ψ µµ = (5.88 ± 0.10)% [16]. Only the fraction η of µ + µ − pairs that satisfies the kinematical conditions 0 < y < 1, (11) −1/2 < cos θ < 1/2 (12) can be registered by the NA50 spectrometer. Here y stands for the rapidity of a µ + µ − pair in the center-of-mass frame of colliding nuclei. θ is the polar angle of the muon momentum in the rest frame of the pair. An estimate of η is impossible without detailed information about the hydrodynamic expansion of the system and the conditions at the thermal freeze-out. We shall therefore treat η as one more free parameter.
In the NA50 experiment the Drell-Yan muon pair multiplicity (either measured or calculated from the minimum bias data) is used as a reference for the J/ψ suppression pattern. Similarly to cc pairs, the number Drell-Yan pairs is proportional to the number of primary nucleon-nucleon collisions: where σ N N DY ′ is the nucleon-nucleon production cross section of µ + µ − Drell-Yan pairs. The prime means that the pairs should satisfy the kinematical conditions of the NA50 spectrometer (11) and (12). As the Drell-Yan cross section is isospin dependent, an average value is used: For the case of Pb+Pb collisions, A = B = 208 and σ P bP b DY ′ = 1.49 ± 0.13 µb [5].
Hence, the quantity to be studied is the ratio It is convenient to rewrite the last expression in a simpler form and treat C and σ N N cc as free parameters. In this form our fitting procedure does not depend on chemical freeze-out conditions. The new free parameter C is connected to η by the expression Here we have introduced the total open (anti)charm density: n O = N O /V and total J/ψ "density": n tot J/ψ = N tot J/ψ /V . The relation between C and η does depend on freeze-out conditions, but our calculations with the parameter sets (8) and (9) have shown that this dependence is not essential.
In the NA50 experiment, the neutral transverse energy of produced particles E T was used to measure centrality of the collisions. This variable, however, provides a reliable measure of the centrality only if it does not exceed a certain maximum value: E T < ∼ 100 GeV (see also Ref. [21,22]). To show this we have calculated the dependence of the average number of participants on the transverse energy N p (E T ).
The conditional probability to measure some value of E T at fixed impact parameter b is given by a gaussian distribution: Analyzing the experimental situation, we are interested in a quite opposite question: how events with fixed E T are distributed with respect to the centrality. The answer is where P int (b) stands for the probability (see Appendix) that two nuclei at fixed impact parameter b interact (at least one pair of nucleons collides). The average number of participating nucleons at fixed E T is then given by the expression: The parameter values q = 0.274 GeV and a = 1.27 [23] are fixed from the minimum bias transverse energy distribution.
The result is shown in Fig.2. As is seen, the transverse energy is simply related to the number of participants E T = qN p in the domain E T < ∼ 100 GeV. Outside of this domain N p does not change essentially as E T grows. Therefore the data at E T > 100 GeV do not represent centrality dependence of the J/ψ suppression pattern but rather its dependence on fluctuations of the stopping energy at fixed number of participants. In principle, influence of such fluctuations on J/ψ multiplicity can be studied in the framework of our model, but information concerning the corresponding fluctuations of the chemical freeze-out parameters T and µ B would be needed. Experimental data that would allow to extract this information (hadron yields at extremely large transverse energy) are not available at present. Therefore, we restrict our analysis to centrality dependence of J/ψ production and do not use the data corresponding to large transverse energies E T > 100 GeV.
On the other hand, the SCM is not expected to describe small systems. This can be seen from ψ ′ data [10]. In the framework of SCM the multiplicity of ψ ′ is given by the formula (3) with the replacement N tot J/ψ → N ψ ′ . Therefore, the ψ ′ to J/ψ ratio as a function of centrality should be constant and equal to its thermal equilibrium value. The experimental data [24] (see also a compilation in Ref. [10]) are consistent with this picture only at rather large (N p > ∼ 100) numbers of participants [25]. Hence, the applicability domain of the model is limited to Note that the most precise and abundant NA50 data (see Fig. 1) correspond to this kinematical region. At E T < ∼ 100 GeV the formula (16) and the equation give a parametric dependence of the ratio R on the transverse energy. This dependence for the parameter set C = (2.59 ± 0.25) · 10 3 σ N N cc = (34 ± 10) µb (23) is plotted in Fig. 1. The free parameters were fixed by fitting three sets of NA50 data [6,7] within the applicability domain (21) of the model by the least square method. The model demonstrates excellent agreement with the fitted data (χ 2 /dof = 1.2). Extrapolation of the fit to peripheral collisions reveals sharp increase of the ratio (15) with decreasing N p . Such behavior in the SCM can be understood as the following. The smaller is the volume of the system the larger is the probability that c andc meet each other at hadronization stage and form a hidden charm meson. As is seen from Fig. 1, this is not supported by the data: the SCM curve lies above the experimental points in the low E T region. On the other hand, the normal nuclear suppression model also fails to explain the leftmost point from the 1996 standard analysis set and two leftmost points from the 1996 minimum bias set. Those theoretical calculations underestimate the experimental values. It is natural to assume that an intermediate situation takes place 2 . Some fraction of peripheral Pb+Pb collisions result in formation of deconfined medium. In these collisions charmonia are formed at the hadronization stage, and their multiplicities are given by SCM. The rest collisions (we shall call them 'normal collisions') do not lead to color deconfinement, therefore charmonia are formed exclusively at the initial stage and then suffer normal nuclear suppression. The experiment measures the average value, which lies between the two curves.
The fraction of 'normal' events decreases with growing centrality. Their influence on J/ψ production becomes negligible at N p > ∼ 100. To check this we repeated the above fitting procedure using only the experimental data corresponding to N p > 200. The quality of the fit is only slightly better: χ 2 /dof = 1.1, the parameter values C = (2.73 ± 0.40) · 10 3 and σ N N cc = (31 ± 12) µb are consistent with the analysis of the full data set (23). Our picture is also supported by ψ ′ data. The normal nuclear suppression influence nascent charmonia before the formation of meson states. Therefore its effect on ψ ′ is the same as on J/ψ. The multiplicity ratio of ψ ′ to J/ψ in 'normal' nuclear-nuclear collisions should be the same as in nucleon-nucleon collisions and should not depend on the centrality. In the framework of SCM, the ψ ′ to J/ψ ratio, as was explained above, should be equal to its thermal equilibrium value, which is a few times smaller than the corresponding value for 'normal' collisions. As the fraction of 'normal' events decreases, the measured ratio should decrease and then become constant and equal to its thermal value. The experimental data [24] indeed demonstrate such behavior [25,10].
The present analysis predicts strong enhancement of the total number of charm. From a pQCD fit of available data on charm production in p+N and p+A collisions, one could expect σ N N cc ≈ 5.5 µb at √ s = 17.3 GeV. Our result (23) is larger by a factor of 4.5 ÷ 8.0, which is around the upper bound of the charm enhancement estimated in Ref. [15].
Formation of deconfinement medium can change not only the total number of charmonia and open charm particles but also their rapidity distributions. For direct charmonium production in hard parton collisions, dimuon pairs satisfying the kinematical conditions (11) and (12) account for a fraction of about η hard ≈ 0.24 in the total number of pairs originating from J/ψ decays. (The value was found using Schuler's parameterization [27].) Our result (23) corresponds to η ≈ 0.14, which is by a factor of about 0.6 smaller. This difference can be attributed to broadening of the J/ψ rapidity distribution. It is natural to expect similar modification of the open charm rapidity distribution. Because of this modification the open charm enhancement within a limited rapidity window can, in general, differ from the one for the total phase space. Assuming that the broadening for the open charm is approximately the same as that for J/ψ, one obtains open charm enhancement by a factor of about 2.5÷4.5 within the rapidity window (11), which is consistent with the indirect experimental result [28] 3 .
In conclusion, we have shown that the NA50 data on centrality dependence of the J/ψ and ψ ′ production in Pb+Pb collisions [6,7,24] are consistent with the following scenario: The deconfined medium, which is formed in a Pb+Pb collision, prevents formation of charmonia at the initial stage of the reaction. Instead, hidden charm mesons are created at the hadronization stage due to coalescence of created earlier c andc quarks. Within this scenario, the color deconfinement does not necessary lead to suppression of J/ψ. Both suppression and enhancement are possible [30]. If the number of nucleon participants is not too small (N p > ∼ 100), the number of produced J/ψ is smaller than in the case of normal nuclear suppression, therefore anomalous suppression is observed. As color deconfinement is present in most collision events for N p > ∼ 100, our model reveals excellent agreement with the experimental data in this centrality domain. The statistical coalescence model does not describe the NA50 data for the peripheral Pb+Pb collisions. It seems that the fraction of events producing the deconfinement medium is not dominating there and most of peripheral collisions follow the normal nuclear suppression scenario. Still, the presence a fraction of abnormal events could reveal itself in the deviation of the J/ψ data up from the normal nuclear suppression curve.
Our model analysis predicts rather strong enhancement of the open charm. This effect can also be related to the color deconfinement [15]. The enhancement within the rapidity window 0 < y < 1 is consistent with the indirect NA50 data [28]. A direct measurement of the open charm would be very important for checking the above scenario. 4π ∞ 0 drr 2 ρ(r) = 1. (A2) The nuclear thickness distribution T A (b) is given by the formula and the nuclear overlap function is defined as From Eq.(A2), one can deduce that the above functions satisfy the following normalization conditions: In Glauber's approach the average number of participants ('wounded nucleons') in A+B collisions at impact parameter b is given by [32] Here σ inel N N is the nucleon-nucleon total inelastic cross section. At large impact parameter, the nuclei may do not interact at all. ThereforeÑ p (b) → 0 at b → ∞. If one interested in the average number of participants, provided that an interaction between two nuclei has taken place, the relevant quantity is where is the probability for nuclei A and B to interact at impact parameter b. Although N p (b) differ fromÑ p (b) at large b:Ñ p (b) → 2 at b → ∞, they are almost identical for more central collisions.
The average number of nucleon-nucleon collisions can be calculated from Provided that an interaction between two nuclei has taken place, the above formula should be modified as | 4,377.8 | 2001-10-20T00:00:00.000 | [
"Physics"
] |
Intelligent personal assistants and L2 pronunciation development: focus on English past -ed
. This study investigates an Intelligent Personal Assistant’s (IPA) ability to assist English as a Second Language (ESL) learners in developing their phonological awareness, perception, and production of the allomorphy in regular past tense marking in English (e.g. talk[t], play[d] and add[ɪd]). The study addresses the following questions: Can the pedagogical use of IPAs improve learners’ pronunciation of -ed allomorphy in terms of phonological awareness, perception, and production? What are learners’ attitudes toward IPAs? The results suggest that participants improved in their ability to articulate their phonological awareness regarding the target form, and that their attitudes toward the technology was positive in terms of the four measures adopted to assess their experience (i.e. learnability, usability, motivation, and willingness to use). We discuss these findings and emphasize the pedagogical potential of IPAs for the development of L2 pronunciation, as well as their ability to personalize learning and consequently extend the reach of the language classroom.
Introduction
The use of technology in language learning provides learners with increased autonomy and opportunities to regulate their own learning, while offering easy access to information outside the language classroom (Braul, 2006). This study explores the pedagogical use of one such technology: IPAs, voice-controlled services that complete tasks by orally interacting with users. An example of a popular IPA is the Alexa App (Alexa henceforth), a virtual assistant developed by Amazon. This study examines whether Alexa can assist English learners in developing and/ or improving their phonological awareness, aural perception, and oral production of the allomorphy that characterizes regular past tense marking in English (e.g. -ed is pronounced talk[t], play [d] or add [ɪd]). The rationale for using IPAs is based on previous research with IPAs (Dizon, 2017;Moussalli & Cardoso, 2016, 2020Underwood, 2017) indicating that their use encourages repetition, improves listening and speaking, and can motivate learners to reformulate, self-correct, and persist in using the L2 (Moussalli & Cardoso, 2020).
In its design, this study adopts Celce-Murcia, Brinton, Goodwin, and Griner (2010) recommendation for pronunciation instruction. The process starts with awareness raising (Phase 1), then with the development of perception or discrimination abilities (Phase 2), and controlled (Phase 3) and guided oral production (Phase 4), toward a more spontaneous and automatized use of the target feature (Phase 5). This study focuses on the first four stages. In addition, it examines the participants' attitudes toward the use of IPAs to assess the tool's potential to promote learning (learnability), its usability, and potential to increase motivation and willingness to use the technology. The following questions guided this study.
• Will the pedagogical use of an IPA (Alexa) assist in the learning of English past -ed allomorphy in terms of phonological/sound awareness, perception (or phonemic discrimination), and production?
• What are learners' attitudes toward the pedagogical use of Alexa?
Method
Eighteen ESL students (nine males, nine females) from different language backgrounds (CEFR 3 scale: B1-C2) were divided into two groups: the Alexa and the non-Alexa group. This study consisted of five main phases: (1) pre-test, (2) explicit -ed instruction, (3) app familiarization (for the Alexa group only), (4) practice, and (5) (2) two perception tests: while one assessed the participants ability to discriminate the three allomorphs in sentences, the other assessed the target allomorphs in words produced in isolation; and (3) two oral production tests: read-aloud tasks for controlled production, and a role-playing game for guided speech. All participants underwent the five phases: pre-testing, explicit -ed instruction, and the practice phase (for both Alexa and non-Alexa group). In addition, the Alexa group was provided with an app familiarization phase. After the four-week pedagogical intervention, the participants completed the post-tests (modified versions of the pre-tests). At the end of the experiment, participants in the Alexa group were asked to complete a survey (using a nine-point Likert scale: 1=strongly disagree, 9=strongly agree) inquiring about their attitudes toward Alexa (learnability, motivation, willingness to use, and usability). They were also interviewed about their experience and attitudes toward the pedagogical use of Alexa to learn about English pronunciation. The participants in the non-Alexa group, on the other hand, were interviewed about their experience completing the assigned learning materials.
A mixed method design was used. Means and standard deviations were calculated for the five-item awareness test and nine-point Likert scale survey (n=28). Between and within mixed ANOVAs were calculated for both sets of perception and production tests. Finally, the interview data were transcribed and analyzed according to the coding methods proposed by Saldaña (2009).
Results and discussion
To answer the first research question, we examined the participants' development across the three levels of testing: awareness, perception, and production. The results for the phonological awareness test revealed that participants improved between the pre-and post-test for the first test (survey). For example, for '-ed accurate' statements (e.g. "-ed in kissed and jumped sounds the same"), means increased from M=5.56 to M=6.61, while for '-ed inaccurate' statements (e.g. "-ed is pronounced the same in walked, lived, and invited") means decreased from M=3.33 to M=2.56, as hypothesized (see Table 1). The results for the two perception tests revealed that there were no significant differences between the pre-and post-test for all measures (Figure 2). Similarly, also shown in Figure 2, the results for production were not deemed significant for any of the tests performed. Interestingly, the results of our qualitative analysis based on interviews revealed that participants in the Alexa group found it easier to produce the target [t], [d], and [ɪd] allomorphs than to perceive them.
Figure 2. Perception and production tests: results
Regarding the second question, the results revealed that Alexa has great potential as a learning tool (learnability: M=7.07 /9), it has high usability scores (usability: M=6.77 /9), it motivates the participants to learn and explore the language (motivation (M=7.3 /9), and it sparks their willingness to continue to use it in their future language learning endeavors (willingness to use: M=7.73 /9, see Table 2, where Cronbach's Alpha values indicate satisfactory internal consistency between the items for each theme adopted). The participants also explained that the IPA was a great tool for use outside the language classroom, as a conversational partner (e.g. "is important because you sometimes don't have other person for speak so Alexa is a tool for this when you are alone [sic]"). In sum, the results revealed some improvements based on the phonological awareness tests but no significant differences between the pre-and post-tests for the perception and production, probably because of the study's limitations: the short duration of the treatment and the low number of participants. However, the participants did find Alexa a great tool for learning and motivating them to use the L2, as attested in quantitative (phonological awareness tests) and qualitative data (interviews), thus corroborating findings highlighting the potential of IPAs to support L2 development (Dizon, 2020;Moussalli & Cardoso, 2020).
Conclusions
This study contributes to the computer assisted language learning literature by demonstrating that IPAs are valuable pedagogical tools that can extend the reach of the classroom by allowing language learners to autonomously improve aspects of their L2 phonological development (e.g. awareness of past tense marking). As far as L2 pronunciation is concerned, this study adds to the existing literature that explores the link between listening (perceptual) training and output practice on the acquisition of L2 morphophonemics. | 1,703.8 | 2021-12-13T00:00:00.000 | [
"Linguistics",
"Computer Science",
"Education"
] |
An Improved Adam Optimization Algorithm Combining Adaptive Coefficients and Composite Gradients Based on Randomized Block Coordinate Descent
An improved Adam optimization algorithm combining adaptive coefficients and composite gradients based on randomized block coordinate descent is proposed to address issues of the Adam algorithm such as slow convergence, the tendency to miss the global optimal solution, and the ineffectiveness of processing high-dimensional vectors. The adaptive coefficient is used to adjust the gradient deviation value and correct the search direction firstly. Then, the predicted gradient is introduced, and the current gradient and the first-order momentum are combined to form a composite gradient to improve the global optimization ability. Finally, the random block coordinate method is used to determine the gradient update mode, which reduces the computational overhead. Simulation experiments on two standard datasets for classification show that the convergence speed and accuracy of the proposed algorithm are higher than those of the six gradient descent methods, and the CPU and memory utilization are significantly reduced. In addition, based on logging data, the BP neural networks optimized by six algorithms, respectively, are used to predict reservoir porosity. Results show that the proposed method has lower system overhead, higher accuracy, and stronger stability, and the absolute error of more than 86% data is within 0.1%, which further verifies its effectiveness.
Introduction
Te introduction of this study is described in the following sections.
Background.
With the rapid development of artifcial intelligence, population optimization algorithms [1], the memetic algorithm [2], and frst-order optimization methods, such as random gradient descent [3] and gradient descent with momentum (SGDM) [4], have been widely used in the feld of machine learning and play an important role in solving optimization problems of complex systems. As a frst-order adaptive step stochastic gradient optimizer, the Adam algorithm has gained a lot of attention in the feld of numerical optimization for its outstanding computational efciency and has been widely used in deep learning with impressive results [5]. However, the frst-order momentum of the Adam algorithm is an exponentially weighted average of the historical gradients, and the update of the search direction is infuenced by the deviation value of the gradient, which leads to slow convergence of the model. While the second-order momentum is accumulated over a fxed time window, and the data do not vary monotonically with the time window. Tis generates oscillations in the learning rate in the later stages of training and leads to failure of the model convergence. Terefore, it has become a focus of researchers to seek methods to improve the defects of the Adam algorithm in convergence. studies focus on further improving the performance of the optimizer or combining it with other optimization methods [6]. By assigning a "long-term memory" to the historical gradients, the AMSGrad [7] algorithm is proposed, which solves the convergence problem theoretically. Based on the momentum-accelerated stochastic gradient descent, Ma and Yarats [8] proposed a quasi-hyperbolic weight decay acceleration algorithm and adjusted the hyperparameters. Luo et al. [9] compared the generalization and convergence capabilities of stochastic gradient descent (SGD) and adaptive methods and provided new variants of Adam and AMSGrad, identifed as AdamBound and AMSBound, respectively, by using dynamic learning rate variation bounds to achieve an asymptotic and smooth transition from adaptive methods to SGD. Yin et al. [10] proposed a C-Adam algorithm based on the current gradient, predicted gradient, and historical momentum gradient to attain iteratively more accurate search directions by updating the true gradient. Subsequently, a hybrid Adam-based optimization method HyAdamC [11] is proposed, which carefully tunes the search intensity using three-speed control functions: initial, short term, and long term, thus, signifcantly enhancing the prediction accuracy. Later, some methods were proposed such as AdaGrad [12], Yogi [13], Fromage [14], difGrad [15], RBC-Adam [16], and TAdam [17].
Although the above optimization algorithms can achieve competent results when used to train neural networks, they still pose the following three problems. First, the algorithms need to determine an optimal search speed at each training step, which may introduce overftting or afect the training accuracy and testing accuracy [18]. Secondly, the current momentum used in the Adam is prone to inaccurate search directions because of gradient deviations caused by the outliers [17]. Tirdly, such algorithms have difculty in identifying the current state of the optimized terrain in the solution space spanned by the weights, and therefore, they fail to fnd the approximate optimal weights.
Contribution.
To deal with the above problems, an improved Adam optimization algorithm, combining adaptive coefcients and composite gradients based on randomized block coordinate descent, written ACGB-Adam, is proposed. Te contributions and innovations of this article are summarized as follows. (1) To deal with the problem of slow convergence of the Adam algorithm, adaptive coefcients are used for computing the degree of diference between the frst-order momentum and the current gradient. Tis helps to reduce the degree of infuence of parameters on the deviated gradient caused by the outlier points, improve the proportion of infuence of the parameters on the momentum at the previous moment, avoid the gradient deviation, and enhance the search speed and convergence accuracy. (2) Aiming at the shortcoming that the Adam algorithm tends to miss global optimal solution, the prediction gradient is introduced and combined with the current gradient and the frst-order momentum to form a composite gradient, thus, providing a joint determination of the direction of the iterative optimization. Tis helps to get a more accurate search direction and improve the global search capability, thereby speeding up the search for the global optimal solution. (3) To address the issue of dealing with high-dimensional vectors and the high computational overhead of the Adam algorithm, the randomized block coordinate descent (RBC) is introduced to determine the gradient update mode according to the random variables of the diagonal matrix. Tis ensures that only one block of the gradient needs to be computed in each iteration instead of the entire gradient. Ten, the dynamic balance between the convergence accuracy and the system overhead can be achieved. (4) Combining the above ideas, the ACGB-Adam optimization algorithm is proposed. Te optimization performance of the proposed algorithm is verifed by standard classifcation datasets Mnist and CIFAR-10, which is further applied to BP neural networks and compared with optimization methods based on SGD, AdaGrad, Adam, C-Adam, and RBC-Adam. From the experimental results, it can be concluded that the algorithm proposed in this article has better performance, and its convergence speed, stability, and prediction accuracy are higher than those of the other fve methods.
Adam Algorithm
Te Adam algorithm is explained in the following sections.
Basic Principles.
Te Adam algorithm [19] difers signifcantly from the traditional SGD algorithms. SGD algorithm maintains a single learning rate to update all the weights during training; the AdaGrad algorithm reserves a learning rate for each parameter to improve the performance on sparse gradients; the RMSProp algorithm adaptively reserves a learning rate for each parameter based on the mean of the nearest magnitude of the weight gradient, thereby improving the algorithm's performance on nonstationary problems. Adam algorithm sets independent adaptive learning rates for diferent parameters by computing the frst-order and the second-order momentum estimates of the gradient and gains the advantages of both the AdaGrad and RMSProp algorithms.
Particularly, the Adam algorithm uses not only frst-order momentum to maintain the direction of the historical gradient but also second-order momentum to maintain the adaptive state of the learning rate. Besides, it directly considers a sequential setting where samples are displayed sequentially rather than assuming that a large number of training samples are preavailable. Because of these reasons, the Adam algorithm performs well with high computational efciency and low memory requirements [20]. In recent years, research on the Adam algorithm has fourished, and several variants such as NAdam [21], GAdam [22], AMSGrad [23], Adafactor [24], and Adadelta [25] have been proposed.
Algorithm Flow.
In view of accurately describing the Adam algorithm and its improvement, the relevant parameters involved in this article are described in Table 1. Te pseudocode of the Adam algorithm is shown in Algorithm 1.
2
Computational Intelligence and Neuroscience
Existing Problems.
In deep learning, the Adam algorithm is widely used to solve parameter optimization problems because of its efcient calculation, smaller number of tuning parameters, and high compatibility. However, there are certain shortcomings of this algorithm. Firstly, the model convergence speed is very slow. Te frst-order momentum in the Adam algorithm is the exponentially weighted average of the historical gradient, which controls the update of the optimization direction. It gets easily affected by the gradient deviation value, leading to poor searchability and slow convergence speed of the model. Secondly, it is easy to miss the global optimal solution. Te neural network model often contains a large number of parameters. In a space with extremely high dimensions, the nonconvex objective function often tends to rise and fall, and it is easy to produce the "plateau phenomenon" that causes the training to stop and then miss the global optimal solution.
ACGB-Adam Algorithm
To solve the problems of the Adam algorithm, the ACGB-Adam algorithm is proposed, which is primarily improved from the following three aspects. (1) To address the slow convergence speed of the Adam algorithm, an adaptive coefcient calculation method is adopted to improve the search direction and reduce the infuence of gradient deviation caused by the outliers on the frst-order momentum search direction. (2) In view of the issue that the Adam algorithm is easy to miss the global optimal solution, a composite gradient is formed out of the current gradient and the predicted gradient, which enhances the correctness of the search direction, improves the global optimization ability, and further boosts the search efciency and optimization ability of the algorithm. (3) To reduce the computational cost of the algorithm, the randomized block coordinate descent method is introduced to select variables by modules to calculate the gradient update mode. Tis contributes to reducing the memory and CPU utilization as much as possible on the premise of ensuring the search performance.
Adjust Gradient Deviation with Adaptive Coefcients.
In the Adam algorithm, the gradient deviation caused by outliers has a signifcant impact on the calculation of the frst-order momentum. From the exponential weighted average (EWA), it can be noticed that the frst-order momentum maintains the movement direction of its historical gradient, so the search direction of the next time is determined by the previous frst-order momentum of the current gradient. Subsequently, if the current gradient is far from the global optimal direction, the direction of the frst-order momentum will be further away from the approximate optimum, leading to a serious decline in the search ability. Figure 1 demonstrates the impact of the desired gradient on the frst-order momentum. As highlighted in Figure 1(a), the frst-order momentum at the current time m t is calculated by the EWA between the previous momentum m t−1 and the current gradient g t , and the two constant coefcients β 1 and (1 − β 1 ) are used to obtain the EWA. At this time, the direction of m t shifts to the direction of g t if g t deviates from the desired direction due to the infuence of the outliers. Terefore, the search direction at the next time will also be further away from the approximate global optimum P * , as demonstrated in Figure 1 To improve the slow convergence speed caused by the deviation of the frst-order momentum search direction, it is mandatory to confrm whether the current gradient is the deviation gradient caused by the outliers and reduce its impact as much as possible. So, the ACGB-Adam algorithm computes the diference between m t−1 and g t . If this difference is very large, g t is more likely to afect the search direction at the next moment than the frst-order momentum. In this case, the infuence of the momentum at the previous time m t−1 will be increased according to their diference degree by an adaptive coefcient to reduce the infuence of g t on m t as much as possible. Te outlier gradient adjustment method based on the adaptive coefcient is expressed as where β 1,t is the adaptive coefcient, which is proportional to the diference between m t−1 and g t , namely, In this article, the method in [17] is used to determine the diference ratio, as mentioned in equation (2). In equation (2), q t denotes the similarity between g t and m t as calculated by equation (3), and d represents the vector dimension. Q t−1 is a weighted cumulative sum of q 1 , q 2 ,. . ., q t−1 , as calculated by equation (4):
Combined Predicted Gradient to Form Composite
Gradient. In the Adam algorithm, the frst-order momentum m t is determined by the current gradient g t and the historical frst-order momentum m t−1 . Tis causes the search direction to be excessively dependent on the historical gradient, making it easy to miss the global optimum. Te ACGB-Adam algorithm thus introduces the predicted gradient u t , updates the parameter to be optimized at the next moment by the gradient descent method, and difers it from the historical momentum so that it uses a real gradient update and then merges with the current gradient and the historical frst-order momentum to form a composite gradient. Tis makes it possible to get a more accurate search direction in the next iteration. Figure 2 illustrates the schematic diagram of the frst-order momentum search direction adjustment mechanism integrating adaptive coefcients and composite gradient.
For the frst-order momentum before improvement in Figure 2(a), a constant coefcient β 1 is used. Terefore, if g t moves away from the optimal position P * in a direction Learning rate β 1 , β 2 Exponential decay rate of the frst-order and second-order moment estimation, respectively T, t Te maximum iterations and the current t time step, respectively Product of exponential decay rate of the frst and second-order moment estimation at t time step, respectively, Te second-order moment vector at t time step g t Current gradient at t time step Te parameter that needs to be optimized f t Te sequence of the smooth convex loss function P * Global optimal position Get a stochastic gradient objective at time step t: Update biased frst-order moment estimation: Get bias-corrected second-order moment estimation End For (10) Return θ t ALGORITHM 1: Adam. 4 Computational Intelligence and Neuroscience deviated from the desired direction, m t will continue its movement in the direction of g t . In Figure 2(b), m 1,t is the search direction corrected by the adaptive coefcient β 1,t . As compared to the direction of m t in Figure 2(a), m 1,t will approach the global optimal position in a more accurate direction. Terefore, to adjust the gradient efect of outliers, an adaptive coefcient is introduced. Because of this, the infuence of the outliers of the frst-order momentum at the previous moment is as small as possible while calculating the current frst-order momentum. Tus, a more potential search direction can be efectively determined, and the search for the global optimal solution can be accelerated. Secondly, based on the use of an adaptive coefcient to correct the search direction, the predicted gradient u t is introduced, and the search direction m t is formed together with the current gradient g t and the historical frst-order momentum m t−1 . It can be observed that, by introducing the predicted gradient, on the basis of the adjustment of m 1,t , the search direction formed can be further closer to P * to avoid missing the global optimal solution. Terefore, the convergence accuracy of the algorithm is improved.
Gradient Update Mode Based on Randomized Block
Coordinate Descent. As a simple and efective method, SGD is often used to learn linear classifers. However, when dealing with high-dimensional vector data, the full gradient descent mode in SGD is not easy to be implemented in parallel. Terefore, this article introduces the random block coordinate method to optimize the Adam algorithm, which can not only handle high-dimensional vectors but also can avoid calculating the complete gradient of all dimensional data in each iteration, thus saving the computing cost and reducing the system overhead on the premise of ensuring the convergence speed and optimization accuracy.
RBC Algorithm.
RBC is a random optimization algorithm [26]. In each iteration, a coordinate (block) is randomly selected, and its variables are updated in the coordinate gradient direction. If f is a convex smooth function and its gradient L i (i ϵ 1, 2, ..., N { }) is a Lipschitz continuous number, the fow of the RBC algorithm is as follows: wherein, x t denotes the parameter vector to be updated. Te RBC algorithm is as shown in Algorithm 2.
RBC algorithm has been widely used to address largescale optimization problems because of its low computation and update cost [16] and its good optimization efect. For instance, Hu and Kwok [27] studied the learning of scalable nonparametric low-rank kernels, and Zhao et al. [28] proposed an accelerated small-batch random block optimization algorithm. Moreover, several machine learning algorithms can be optimized with the help of RBC. For instance, Singh et al. [29] improved the gradient projection algorithm by using RBC, and Xie et al. [30] combined the RBC algorithm with mean-variance optimization.
Gradient Calculation Based on RBC.
In this article, a new gradient calculation method is proposed based on the RBC method. Let D t (t � 0, 1, 2, ..., N) be a n-dimensional diagonal matrix in the tth iteration, and the ith element on the diagonal is denoted as d t i . Here, d t i is a Bernoulli random variable that satisfes the independent identically distributed, i.e., d t Select coordinates iϵ 1, 2, ..., N { } evenly and randomly (5) Update Input: α, β 1 , β 2 , f t Output: θ t (1) Initialize parameters (adaptive coefcient β 1,t , predicted gradient u t , and the remaining parameters were initialized in the same way as in the Algorithm 1) Generate a random diagonal matrix D t / * Gradient Calculation based on Algorithm 2-RBC * / (4) Get a stochastic gradient at time step t: Update the parameters according to the gradient descent method: Get a predicted stochastic gradient at time step t: u t � ∇ θ f t (θ t−1 )/ * Optimization of the frst moment estimation * / (7) Update biased frst-order moment estimation: Compute bias-corrected frst-order moment estimation: Compute bias-corrected second-order moment estimation End For (13) Return θ t ALGORITHM 3: ACGB-Adam.
Te RBC method is used to randomly select a block (subset) from the whole element of a high-dimensional vector through equation (5) If d t i � 1, which means that the corresponding coordinates are selected, then the ACGB-Adam algorithm is executed for gradient calculation; if d t i � 0, which means that the corresponding coordinates are not selected, then the gradient update calculation is not performed. Tus, in each round of gradient updating, only one block (subset) of the gradient has to be computed, and the frst-order and secondorder momentum are calculated based on this. Moreover, it is not necessary to calculate the entire gradient. Terefore, compared with the other full gradient descent algorithms, the optimization method based on randomized block coordinate descent may save a lot of computing costs and reduce CPU utilization as well as memory utilization while ensuring the convergence of the algorithm. Te specifc calculation process is shown in Figure 3.
ACGB-Adam Algorithm Process.
Te ACGB-Adam algorithm process is described in the following sections.
Overall Architecture of Algorithm.
Te overall architecture of the ACGB-Adam algorithm is shown in Figure 4, which mainly includes three core modules: the random block coordinate method, the adjustment of gradient deviation values through adaptive parameters, and the composite gradient.
Te general strategy of the ACGB-Adam algorithm is to integrate the above three modules and apply three optimization methods to solve problems in parameter updating so as to improve the convergence speed, global optimization ability and reduce the system overhead. First, the current gradient update mode is optimized by RBC, which can avoid calculating all gradients and reduce the system overhead. Secondly, through the adaptive parameters, the algorithm could calculate the coefcient proportion of the frst momentum adaptively according to the diference between the current gradient and the frst momentum at the last time so as to minimize the infuence of the outlier gradient and optimize the search direction and search speed. Finally, the composite gradient combines the predicted gradient, the current gradient, and the frst momentum of the last time to form the fnal search direction, aiming to further approach the global optimal position and improve the global search ability of the algorithm.
ACGB-Adam Algorithm Process.
Te overall algorithm fow of ACGB-Adam is shown in Algorithm 3.
Experiment and Analysis
Te experiment and analysis are described in the following sections.
Standard Datasets and Experimental Setup.
To evaluate the performance of the ACGB-Adam algorithm, experiments were carried out on two standard datasets (Table 2) used for classifcation. Te proposed algorithm was further compared with the stochastic gradient descent (SGD), the adaptive gradient (AdaGrad), the adaptive moment estimate (Adam), the Adam optimization algorithm based on adaptive coefcients (A-Adam), Adam optimization algorithm based on composite gradient (C-Adam), and Adam optimization algorithm based on randomized block coordinate descent (RBC-Adam) algorithms ( Figure 5(a)).
(1) Mnist Dataset. Te Mnist dataset [31] developed by the US postal system is a classic dataset for image recognition. In this dataset, 70000 digital pictures of 0∼9 handwritten by 250 diferent people are counted. Tese numbers have been standardized in size and are located in the center of the image. Some examples of handwriting in the dataset are represented in Figure 5(a). (2) CIFAR-10 Dataset. Te CIFAR-10 dataset [32] is used for identifying universal objects which consists of 60000 RGB images. Compared with the handwritten characters, this dataset contains pictures of real objects in the real world. Te noise is large, and the proportions and characteristics of objects are diferent, which lead to great difculties in recognition. Figure 5(b) lists ten classes in the dataset, and each class shows ten pictures randomly. (3) Experimental Setting. MATLAB is used for the simulation of experiments. Te operating system is Win10, the CPU is Intel i7-1065G7, the primary frequency is 1.30 GHz, the memory is 16 GB, and the SSD capacity is 512 GB. To improve the comparability of the results, the six comparison algorithms involved in the experiment all use the same parameter settings. Te main superparameters are as follows: α � 0.001, β 1 � β 2 � 0.9, and the maximum number of iterations is 100. MSE and accuracy are used as performance evaluation indicators of algorithm training and classifcation accuracy.
Experimental Results of the Standard Dataset.
Te experimental results of the standard dataset are explained in the following sections. Figure 6 represents the training error loss and classifcation accuracy of the six algorithms on the Mnist. Te training error and test accuracy at the 100th iteration are shown in Table 3. It can be observed from Figure 6 and Table 3 that, as the number of iterations increases, each algorithm gradually converges on the Figure 7 demonstrates the training error of algorithms on the CIFAR-10 dataset, along with the classifcation accuracy of the test set.
Mnist Experimental Results.
Te training error and test accuracy at the 100th iteration are shown in Table 4. It can be observed from Figure 7 and Table 4 that the training error of the ACGB-Adam algorithm in the early stage of iterations reduces quickly and gradually tends to be stable. With the increase in iterations, the error loss of the algorithm still decreases steadily. Compared with the other six algorithms, the ACGB-Adam algorithm has the smallest error loss value and the highest classifcation accuracy of 0.941. From the experimental results on the CIFAR-10 dataset, it can be inferred that the proposed algorithm in this article has better optimization performance than the other six algorithms in terms of convergence speed, accuracy, stability, and classifcation accuracy.
Memory and CPU Usage Rate Analysis.
For the two standard datasets, the changes in memory and CPU utilization of the seven algorithms with the number of iterations are illustrated in Figure 8 and Table 5. It can be observed from Table 5 and Figure 8 that, with the increase of iteration times, the memory and CPU Computational Intelligence and Neuroscience utilization of each algorithm increase gradually. Under the same conditions, the memory and CPU utilization of the RBC-Adam algorithm is the lowest, followed by the ACGB-Adam algorithm proposed in this article. Te diference between the memory and CPU utilization rates of the two algorithms is less than 2%. Te specifc experimental results are shown in Table 6. It can be seen from Tables 5 and 6 that although the computing cost of the RBC-Adam algorithm is slightly lower than the ACGB-Adam algorithm, its training error and classifcation accuracy are far lower than those of the proposed algorithm. Altogether, the ACGB-Adam algorithm proposed in this article can achieve a dynamic balance in convergence and computing cost. On the premise of improving the convergence speed and accuracy, it can reduce the memory and CPU utilization to the greatest extent and has good comprehensive optimization performance.
Reservoir Porosity Prediction.
To further verify the effectiveness and utility of the algorithm proposed, the reservoir porosity in the real work area was predicted by a BP neural network based on the ACGB-Adam algorithm. Figure 9, the sample data are from the real data of two wells, A and B, in an exploration area. Te logging depth is 900∼1120 m, including 1492 records and 11 logging parameters. To achieve efcient and accurate porosity prediction, the grey correlation analysis method [33] is used to select parameters with high correlation with porosity as input parameters of the neural network, namely, Depth, RLLS (shallow investigate double lateral resistivity log), GR (natural gamma ray), HAC (high-resolution interval transit time), and DEN (density), as represented in Figure 10. Tis helps to improve the data processing efciency on the premise of ensuring prediction accuracy. It can be assumed that these fve parameters that have a signifcant impact on porosity are diferent in nature and usually have distinct dimensions and orders of magnitude. In case the level diference between the parameters is too large, the infuence of the parameters with higher values will be highlighted, and the efect of the parameters with lower values will be weakened. To ensure the comparability of the data, this article uses the deviation normalization method [33] to preprocess the data and eliminates the infuence of the dimension and the value of the variable itself on the results.
Model Performance Analysis.
Te preprocessed data were taken as sample data, and the training set and test set were divided in the ratio of 8 : 2. Te BPNN model is set as follows: the number of hidden layers was 1, including 5 neurons, the transfer function was Tansig, the learning rate was 0.001, and the maximum number of iterations was 5000. Using MSE and RMSE as the model performance evaluation indices, the proposed ACGB-Adam_BP model was compared with fve methods, namely, SGD_BP, AdaGrad_BP, Adam_BP, C-Adam_BP, and RBC-Adam_BP. Te fnal training error and test error of various methods are enlisted in Table 7, in which the minimum values of MSE and RMSE are shown in bold, and the iterative error curve is shown in Figure 11.
It can be seen from Table 7 and Figure 11 that the BPNN based on the ACGB-Adam algorithm generates the lowest error in the training set and the test set and tends to be stable as soon as possible. Te convergence speed is much better than the other fve comparison algorithms. Tis indicates that the proposed algorithm has better optimization performance.
Porosity Prediction Results.
To further observe the above results intuitively and validate the efectiveness and correctness of the method proposed in this article for porosity prediction, the prediction results of the BP model based on the ACGB-Adam optimization algorithm are visually analyzed in terms of 300 test samples, as highlighted in Figure 12. Due to space constraints, the error analysis results on the training set are not shown in the article. From the comparison curve between the predicted value and the actual value of porosity, it can be observed that the BP neural network model based on the ACGB-Adam optimization algorithm has a relatively ideal prediction result, and the predicted abnormal value of porosity is quite less. Te absolute error of more than 86% of the data is Table 7 is shown in bold.
Computational Intelligence and Neuroscience within 0.1%, which signifes the high prediction accuracy of the proposed algorithm.
Conclusion
Starting with the improvement of the Adam algorithm to heighten the convergence speed, accelerating the search for the global optimal solution, and enhancing the high-dimensional data processing ability, the Adam optimization algorithm combining adaptive coefcients and composite gradients based on randomized block coordinate descent is proposed, which enhances the performance of the algorithm. Trough theoretical analysis and numerical experiments, the following conclusions can be drawn: (1) Te gradient deviation caused by the outliers is crucial to the convergence speed and solution precision of the Adam algorithm. Using an adaptive coefcient to adjust the diference between the frstorder momentum and the current gradient can help in reducing the infuence of parameter proportion of deviation gradient, improving the slow convergence speed of the Adam algorithm, boosting the search speed, and improving the convergence accuracy. (2) By introducing the prediction gradient and combining the current gradient and the frst-order momentum to form a composite gradient, an accurate search direction can be obtained in the subsequent iteration, and then, the global optimization ability of the algorithm could be enhanced. (3) In the process of gradient updating, the RBC method is used to determine the gradient calculation method by randomly selecting variables from the parameter subset. Tis can reduce the calculation cost as much as possible on the premise of ensuring the convergence of the algorithm, enhance the processing ability of the algorithm for high-dimensional data, and maintain a good balance between the optimization accuracy and the system overhead. (4) Te test results on Mnist and CIFAR-10 standard datasets for classifcation indicate that the ACGB-Adam algorithm is signifcantly superior to SGD, AdaGrad, Adam, A-Adam, C-Adam, and RBC-Adam algorithms in terms of convergence speed and optimization accuracy. Although the proposed method is slightly higher than the RBC-Adam algorithm in terms of memory and CPU utilization, it can achieve a decent balance between convergence and system overhead.
According to the evaluation indices, the proposed algorithm has better performance advantages compared with the other fve algorithms, which validates the efectiveness of the algorithm improvement. (5) Te BPNN model based on the ACGB-Adam algorithm is applied to reservoir porosity prediction. Te experimental results suggest that, as compared to the BPNN model based on Adam and its variants, the maximum reduction of MSE and RMSE of the proposed model in this article is approximately 86.30% and 62.99%, respectively, which achieves higher accuracy in porosity prediction, verifes the superiority of the proposed algorithm, and extends the application feld of the algorithm.
Te method proposed in this article enhances the performance of the Adam optimization algorithm to a certain extent, but does not consider the impact of the second-order momentum and diferent learning rates on the performance of the original algorithm. Terefore, the follow-up research can focus on the optimization and improvement of the second-order momentum and learning rate and conduct indepth and detailed research on the parts not involved in this algorithm. Tis can help to attain better optimization performance.
Data Availability
No data were used to support the fndings of the study.
Conflicts of Interest
Te authors declare that they have no conficts of interest. | 7,429.8 | 2023-01-10T00:00:00.000 | [
"Computer Science"
] |
Experimental dataset from a central composite design with two qualitative independent variables to develop high strength mortars with self-compacting properties
Fresh and hardening properties of cement-based materials are key factors for correctly choosing the constituent materials and their mix proportions. To optimize design-based mortar compositions for specific applications, response models are frequently applied to data collected from scientific approaches. Here, experimental dataset regarding to a design of experiments carried out in mortars through a central composite design with five independent variables is presented. Among the five independent variables, four were quantitative ones: Waterv/Cementv, Superplasticyzerm/Powderv, Waterv/Powderv, Sandv/Mortarv. The other independent variable was a qualitative one: Superplasticiser A or Superplasticiser B. In total 60 mortar compositions were done: for each qualitative variable a 24 factorial design comprising of 16 treatment combinations enlarged by 8 axial runs plus 6 central runs, resulting in a central composite design with 30 mortar trial mix compositions. The following dependent variables were tested: the D-flow and the t-funnel to evaluate the fresh properties and the compressive at the age of 24 h and at the age of 28 days to evaluate the hardened properties. Based on this dataset, response models can be applied to find optimized mix compositions, with the effect of the two qualitative variables being determined.
a b s t r a c t
Fresh and hardening properties of cement-based materials are key factors for correctly choosing the constituent materials and their mix proportions. To optimize design-based mortar compositions for specific applications, response models are frequently applied to data collected from scientific approaches. Here, experimental dataset regarding to a design of experiments carried out in mortars through a central composite design with five independent variables is presented. Among the five independent variables, four were quantitative ones: Water v /Cement v , Superplasticyzer m /Powder v , Water v /Powder v , Sand v /Mortar v . The other independent variable was a qualitative one: Superplasticiser A or Superplasticiser B. In total 60 mortar compositions were done: for each qualitative variable a 2 4 factorial design comprising of 16 treatment combinations enlarged by 8 axial runs plus 6 central runs, resulting in a central composite design with 30 mortar trial mix compositions. The following dependent variables were tested: the D-flow and the t-funnel to evaluate the fresh properties and the compressive at the age of 24 h and at the age of 28 days to evaluate the hardened properties. Based on this dataset, response models can be applied to find optimized mix compositions, with the effect of the two qualitative variables being determined.
Value of the Data
• Typical data published in papers based in a Central Composite Design applied to mortar mix compositions has no more than 20 experiments. The design of experiments is based in three independent variables. Here, the total volume of sand was an independent variable, too. That means an extra independent variable, therefore increasing the number of experiments for 30 experiments. • Besides, two superplasticizers were used as qualitative independent variables. That leads to double the size of the experiments and allows analyses based on qualitative variables which are barely found in the literature. • The mechanical properties are usually measured for one age. Here, the compressive strength was measured at the age of 24 h and at the age of 28 d. Thus, analyses both at early and later ages are possible. • With a total of 60 mix composition testes (30 from each Central Composite Design regarding to a distinct superplasticiser), the outliers/errors analysis and statistical analysis are statistically better verified and surpassed by the large number of experiments. Then, the next coming response models are expected to be more precise.
• Researchers that are working/developing/applying/validating models related to mortars with self-compacting properties and/or with high strength may be interested in this data. • This data may be applied to develop response models to study the effect of the constituent materials in the fresh and hardened properties. Researchers that are dealing with designing approaches of mix compositions with artificial intelligence may be interested in this data, too. Table 1 establishes the equivalence between the coded and the real values. Table 2 displays the dataset for the mix proportions and the corresponding coded values of the mortars prepared for 1.61 L. The proportions regard to a design of experiments based on a central composite design for four independent variables. The dataset base includes 30 mixes. However, due to using qualitative variable -two superplasticizers -a second one central composite design of 30 mixes was added. Note: in Table 2 , the total water added to the mixes was adjusted to the water effective plus the water to saturate the aggregates minus the water included in the superplasticiser.
Experimental Design, Materials and Methods
In this experimental program two powders, two superplasticisers and one sand were used. The powders were the cement CEM I 42.5 R (EN 197-1 [1] ) and the limestone filler, the specific gravity being 3.1 g/cm 3 and 2.68 g/cm 3 , respectively. The sand was the normensand (EN 196-1 [2] ) with specific gravity of 2.63 g/cm 3 and water absorption of 0.30%. The superplasticizers were from third generation, polycarboxylate based, one (A) with ρ= 1.08 g/cm 3 and solid content of 40% and another (B) with ρ= 1.07 g/cm 3 and solid content of 30%. Distilled water was used. Table 1 .
The following procedure was done to produce mixes: (i) the constituent materials were previously weighed and stored in plastic containers; (ii) in a standard mixer for mortars [2] , 80% of the total water was joined to the powders and the sand (that moment was defined as being ( continued on next page ) n.av.: The reading value is not available due to problems in the press equipment. t = 0), and then the constituent materials were mixed at low speed for 120 s; (iii) stop mixing for 60 s to clean the paddle and (iv) add the superplasticiser and the remaining water in the last 10 s; (v) re-started mixing at low speed and mix for 120 s; (vi) stop mixing for 30 s to clean the paddle; (vii) re-started mixing at high speed and mix for 30 s. Just after mixing the V-funnel test and the slump test were performed. From these tests, the reading of the flow time (t-funnel -see Ref. [3] for details) and the readings of the two orthogonal distances of the slump test (D-flow1, D-flow2 -see Ref. [3] for details) were taken. After that, five prismatic specimens 4 × 4 × 16 [cm] were moulded. These five specimens were stored in a climatic chamber at 20 °C.
All the specimens were unmoulded at the age 23.5 h. The compressive strength test was carried out in three specimens at the age of 24 h (with 15 min tolerance) and the others two specimens were stored in a climatic chamber at 20 °C and at least 98% of RH up to the age of 28 d. At the age of 28 d the last two specimens were tested to the compressive strength, too. Note that, previously to the compressive strength test, the tensile strength test [2] was done in the specimens to break the specimens in two halves (no values were recorded). Immediately after that, the halves were tested to the compressive strength [2]
Ethics Statements
Nothing to declare.
CRediT Author Statement
Lino Maia: I did all the research work. No other scientific contributions.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. | 1,782.2 | 2021-12-21T00:00:00.000 | [
"Materials Science"
] |
Recognizing ion ligand binding sites by SMO algorithm
Background In many important life activities, the execution of protein function depends on the interaction between proteins and ligands. As an important protein binding ligand, the identification of the binding site of the ion ligands plays an important role in the study of the protein function. Results In this study, four acid radical ion ligands (NO2−,CO32−,SO42−,PO43−) and ten metal ion ligands (Zn2+,Cu2+,Fe2+,Fe3+,Ca2+,Mg2+,Mn2+,Na+,K+,Co2+) are selected as the research object, and the Sequential minimal optimization (SMO) algorithm based on sequence information was proposed, better prediction results were obtained by 5-fold cross validation. Conclusions An efficient method for predicting ion ligand binding sites was presented.
Introduction
Ions play an important role in the structure and function of proteins: for example, the SO 4 2− participate in the synthesis process of Cysteine [1], the sulfation process after protein translation [2], the synthesis process of proteoglycan, the sulfate absorption and decomposition process of plant and others [3]; the PO 4 3− is an important component of bones and teeth which can maintain the neutrality of body fluids; alkali metal K + and Na + control the charge balance in cells, tissue fluids and blood, which plays an important role in maintaining the normal circulation of body fluids and controlling the acid-base balance in the body; alkaline earth metal Ca 2+ plays a regulatory role in nerve conduction and blood coagulation; transition metal Fe 3+ plays an important role in the oxidative damage process of proteins, lipids, sugars and nucleic acids [4]. The interaction of proteins with ion ligands determines the realization of these biological functions, so the recognition of ion ligand binding sites is important for the study of its function [5][6][7][8][9][10].
In 2002, Richard et al. [11] have tested sulphate ion binding site of proteoglycan, and they identified the sites that is interaction with heparan sulfate. In 2017, Li et al. [12] used protein structural classification (SCOP) and Protein Data Bank (PDB) databases to extract 1251 protein chains using Ligand-Protein Contacts (LPC) software, and gave predictions of 8112 binding residues, and the Support vector machine (SVM) algorithm was used to predict the sulfate ionbinding residues of proteins. In recent years, the Zhang Lab team has compiled a database of ligand-binding residues named as the BioLip [13] database, a semi-manual database that collects interactions between ligands and proteins, functional annotations are relatively comprehensive compared with other databases, which contain extremely extensive and accurate ligand protein data.
During the last few years, many approaches have been developed to predict the binding sites of protein-metal ions. In 2008, Babr et al. [14] predicted the binding sites of protein chains and transition metal ions by CHED algorithm; when predicting 349 whole proteins, 95% specificity was obtained, and 82 prions were predicted to obtain 96% specificity. In 2012, Lu et al. [15] used the "fragment transformation" method to predict metal ion (Ca 2+ , Mg 2+ , Cu 2+ , Fe 3+ , Mn 2+ , Zn 2+ ) ligand binding sites, and the prediction results were obtained with a total accuracy of 94.6% and a true positive of rate 60.5%. In 2016, Hu et al. [16] identified four metal ions in the BioLip database by both sequence-based and template-based methods, and the Matthew's correlation coefficient (MCC) values were greater than 0.5. In 2017, Cao et al. [17] used the SVM algorithm to identify ten metal ion binding sites based on amino acid sequences, which obtained a good result by 5fold cross validation. In 2018, Greenside et al. [18] used an interpretable confidence-rated boosting algorithm to predict protein-ligand interactions with high accuracy from ligand chemical substructures and protein 1D sequence motifs, which got a great result.
In this paper, the dataset of acid radical ion and metal ion ligands was extracted from BioLip database, the Sequential minimal optimization (SMO) algorithm was proposed to predict the binding site with component information, position conservation information and refinement characteristics, experiment results show that the MCC values of the four acid radical ion ligands by 5-fold cross validation exceeded 0.470, the accuracy values were not less than 74.0%; the MCC values of six metal ion ligands of Zn 2+, Cu 2+ , Fe 2+ , Fe 3+ , Mn 2+ and Co 2+ exceeded 0.620, the accuracy values were not less than 80%; the MCC values of four metal ions of Ca 2+ , Mg 2+ , Na + and K + exceeded 0.430, the accuracy values were not less than 71%.
Dataset
The construction of the dataset is directly related to the reliability of the prediction accuracy. The dataset constructed in the paper was from the BioLip database.
The binding protein chains, including four acid radical ion ligands (NO 2− , CO 3 ) and ten metal ion ligands (Zn 2+ , Cu 2+ , Fe 2+ , Fe 3+ , Ca 2+ , Mg 2+ , Mn 2+ , Na + , K + , Co 2+ ), were downloaded from the BioLip database, wherein the sequence length is greater than 50 residues, the resolution is less than 3 Å, and the sequence identity threshold is less than 30%. Then, the sliding window method is adopted to get the overlapping segment on the protein chain, if the center of the segment is the ligand binding site, it is defined as a positive sample; otherwise it is defined as a negative sample. We selected the datasets with the sequence segment length of 17 as an example to simply explain the multiple relationships of segments' number in positive and negative sets; the detailed datasets are summarized in Table 1.
Since the number of samples in negative set is several tens of times the number of samples in positive set, in order to ensure stable of the results, the negative set with equal numbers of positive set was randomly selected ten times in the 5-fold cross validation, and finally the final result was obtained by selecting an average of ten times.
The statistical analysis of dataset Amino acid composition information
According to the literature [12,17], amino acid composition information is an important feature in the recognition of binding sites. Therefore, we analyzed the composition information of acid radical ion and metal ion ligand. The SO 4 2− ligand was taken as an example, the violin plot was shown in Fig. 1. The violin plot is a combination of a box plot and a kernel density, and is mainly used to display the distribution state of the data. The left side of each group represents the amino acid composition in the negative set, the right side represents the amino acid composition in the positive set, the ordinate represents the frequency of occurrence of the amino acid, and the white dot represents the median. The black box pattern ranges from the lower quartile to the upper quartile, representing the concentrated distribution of amino acid; the outer shape represents the kernel density estimation, the more concentrated the data, the fatter the graph. Figure 1 showed that the concentrated distribution interval of R, S and T in the positive set was larger than the concentrated distribution of the negative set, while the D, E, G in the negative set were more concentrated than the positive set. Since the concentrated distribution interval of amino acid composition in the positive and negative sets was significantly different, we used the amino acid composition information as a characteristic parameter.
The position conservation of amino acids
The WEBLOGO [19] software was used to analyze the position conservation of acid radical ion and metal ion ligands. Since the ion ligands are small ligands, they usually only bind with a few residues. So we selected a window length L of 17 as an example to analyze. The x-axis represents 17 positions, the y-axis represents the conservation of amino acids in every position, with the height of each letter corresponding to the occurrence probability of the corresponding amino acid, the center of the positive set indicates the ion ligand binding residue. As shown in Fig.2, the position conservation of the SO 4 2− binding residues and environmental residues are strong, but binding residues are more conservative, the preferred residues are R, G, K, S, H, T, and there is a significant difference of amino acid conservative between positive set and negative set. For example, at the eighth position, the highest frequency of the amino acid is G, S, A, L in positive set; the highest frequency of the amino acid in negative set is L, A, G, V. In the tenth positive, the highest frequency of amino acid is G, T, S, A in positive set; the highest frequency is L, A, G, V in negative set. The above analysis shows that the position conservation of amino acid residues is a good indicator of protein ion binding, so it was selected as the characteristic information to further develop an effective identification model. The selection of characteristic parameters The characteristic parameters from statistical analysis According to the statistical analysis of component information and position conservation information for amino acid, these two kinds of information were selected as characteristic parameters.
Physicochemical properties of amino acids
According to the biological background, the physicochemical properties of amino acid residues play an irreplaceable role in the binding of proteins to ions. Therefore, we chose the hydropathy and polarization charge of amino acids as characteristic parameters. The 20 amino acids are grouped into 6 kinds [20] according to hydropathy characteristic ( Table 2)
Predicted structural information
The prediction of secondary structure and solvent accessibility reflect the spatial structure information of the backbone and side chains [22], so we also extracted these information as characteristic parameters using ANGLOR [23] software. According to the predicted secondary structure information, the 20 amino acids are divided into 3 categories: α-helix, β-sheet and coil; according to the predicted relative solvent accessibility (SA), the 20 amino acids are divided into 2 categories: SA value is greater than 0.25 for exposure; SA value is less than 0.25 for burial.
The extraction of characteristic parameters
According to the statistical analysis, the component information of five characteristic parameters of amino acid, hydropathy, charge, secondary structure and relative solvent accessibility were selected, and the Increment of Diversity algorithm was used to reduce the dimension of the above five components to extract their refinement features; the Position matrix scoring algorithm was used to extract the site information of five characteristic parameters and reduce the dimension to extract their refinement features.
Position matrix scoring algorithm
The Position matrix scoring algorithm constructs a positional frequency matrix using known sequence patterns to describe the composition of amino acids at various positions in an unknown sequence pattern, and to characterize the position conservation of amino acids in the sequence. Through statistical analysis of the ion ligands in this study, it is found that they have obvious position conservation, so the Position matrix scoring algorithm was selected to extract the feature parameters.
Position matrix scoring algorithm is a classification algorithm. It has been successfully used in predicting transcription factor binding sites in genomes and supersecondary structures [24,25].
The position frequency matrix is defined as: In the above equation, j is 20 amino acids and one pseudo amino acid "X", n i, j is the frequency of the j th amino acids at the i th position, N i is total number of all amino acids occurring at the i th position, P i,j is the observed probability of the j th amino acids at the i th position.
The matrix element of the position weight matrix is defined as: P 0,j is background probability of the j th amino acid, m i,j is the weight probability of the j th amino acids at the i th position.
The scoring(S) value is given by the following equation: Here, S is the scoring matrix function, L is length of amino acid sequence segment, C i is conservation index at the ith position, m i,min is the minimum value at the i th position, m i,max is the maximum value at the i th position.
Taking the position amino acid residue as a parameter, two standard scoring matrices were constructed using the training set. In the test set, two scoring (S) values can be obtained for an arbitrary sequence segment, which can be used as the refinement characteristic parameters. Besides, the characteristic parameters of the 2 L dimensional site information can also be obtained by using the position weight matrix.
Increment of diversity (ID) algorithm
Dispersion is a measure of information diversity. It can quantitatively describe certain feature information contained in an amino acid sequence, and the measure of diversity can describe the overall diversity. The increment of diversity is one of the information coefficients. It is applied to the information classification as a classification algorithm. It can reduce the dimension and use the refined features as the characteristic parameters of classification prediction. It has been successfully applied to protein folding and protein structure classification prediction [26,27]. Therefore, the Increment of Diversity algorithm was used to extract the feature information from sequence.
In the state space of dimension S, for a vector X: [n 1, n 2 , …,n s ] the measure of diversity source was For two state spaces of dimension S, for vectors X: [n 1, n 2 , … n s ] and Y: [m 1 , m 2 , …, m s ], the measure of mixed diversity source X + Y was The increment of diversity between the source of diversity X and Y was The amino acid composition information was input into the ID algorithm. The standard discrete source is constructed by training. Two discrete increment (ID) values can be obtained for each segment of the test set. Then, the obtained two-dimensional ID value can be used as the characteristic parameter.
Algorithm
The SMO algorithm was proposed by Platt in 1998, which is also known as the sequence minimum optimization method. It is the fastest quadratic programming optimization algorithm that can effectively improve computational efficiency. The SMO algorithm optimizes only two variables at a time, regards all other variables as constants, transforms a complex optimization problem into a relatively simple two-variable optimization problem, and adopts analytical method to avoid the error accumulation caused by iteration method, which ensures its accuracy. In this paper, we established our identification model using the SMO algorithm based on the weka3.8 [28,29] and using the Precomputed Kernel Matrix (PUK) kernel function. PUK is a general kernel function based on Pearson's seventh function [30]. It has good robustness and has equivalent or even stronger mapping ability than standard kernel functions. It can be used as a general kernel function to replace ordinary linear, polynomial and radial basis kernel functions. To a certain extent, it can eliminate the trouble of how to select the kernel function in the SVM algorithm, saving time.
Performance measure
We used the following four standard measures [31] to evaluate the performance of the identification of ion binding residues: sensitivity (S n ), specificity (S p ), accuracy of prediction (Acc) and Matthew's correlation coefficient (MCC). These were calculated by the following formulae: Where TP is the number of correctly identified acid radical or metal ion binding residues, FN is the number of binding residues wrongly identified as non-binding residues, TN is the number of correctly identified nonbinding residues, and FP is the number of non-binding residues identified as binding residues.
Results and discussion
The optimal window size Whether the amino acid residue can be combined with the ion ligand depends not only on amino acid residue itself but also on neighboring residues [32]. In order to extract more comprehensive information, we used the sliding window method, where different window sizes range from 5 to 17, intercepting the sequence segments from the N-terminal to the C-terminal, and ensuring that all residues appear in the center of the segment, we added an (L-1)/2 dummy residue "X" at both terminals of the proteins. If the central residue of the segment was an ion binding residue, we assigned the segment as positive; otherwise it was assigned as negative. Taking SO 4 2− ligand as an example (Fig. 3), the x-axis represents the window size, the y-axis represents the MCC, ACC, S n and S p values under different window sizes, we performed a large range search on the window size of 7 kinds of amino acid residues and combined the WEBLOGO diagram of the ion ligand to finally determine the optimal window size of SO 4 : 11, 13, 9, 7, 13, 9, 9, 9, 9, 7, 9, 11, 11. The following calculations were made under the optimal window sizes and the 5-fold cross validation commonly used in the literature [33][34][35].
The results under component information parameters
Under the optimal window size, amino acid component information, hydropathy component information, charge component information, secondary structure component information, and relative solvent accessibility component information were collectively used as characteristic parameters and input to the SMO algorithm. The calculation results of 5-fold cross validation were shown in Table 3.
It can be observed from Table 3 that the ACC values of the four acid radical ion ligands were all greater than 61.0%, the MCC values of CO 3 2− , SO 4 2− and PO 4 3− exceed 0.360, and only the MCC value of NO 2 − was lower than 0.225; among the recognition results of metal ion ligands, Zn 2+ , Cu 2+ , Fe 2+ , Fe 3+ and K + were preferable, and the MCC values were not less than 0.5. It can be considered that these five metal ion ligands were sensitive to the component information; the results were consistent with the previous research results. The reason can be seen from the statistical diagram of the amino acid composition given in [17] that the differences of positive and negative sets of transition metal ions were relatively large, so their prediction results were better, and the remaining ion ligands will continue to be identified by adding other characteristic parameters.
The results under position conservation information parameters
Under the optimal window size, we identified the ion ligand binding sites using position amino acid, position hydropathy, position charge, position secondary structure and position relative solvent accessibility as characteristic parameters via the SMO algorithm. The calculation results by 5-fold cross validation were shown in Table 4.
From Table 4, it can be concluded that the MCC value of NO 2 − was 0.350, the MCC value of CO 3 2− was 0.462, the MCC value of SO 4 2− was 0.460, and the MCC value of PO 4 3− was 0.548. Compared with all component information as characteristic parameters, the recognition result has been improved.
For the identification results of ten metal ion ligands, the six metal ion ligands of Zn 2+ , Cu 2+ , Fe 2+ , Fe 3+ , Mn 2+ and Co 2+ have good prediction results, and the MCC values were not less than 0.600; Na + and K + have worst recognition results, we considered that these two ion ligands were less sensitive to the position conservation information and can continue to identify their refinement. Compared with the identification of all the component information as characteristic parameters, the MCC values of Na + and K + decreased slightly, but other's MCC values showed an upward trend, indicating that these ion ligands were more sensitive to the position conservation information, as can be seen from the WEBLOGO in [17]. The positive and negative sets are more different than the statistical analysis of the components in [17], so the ion ligands were more sensitive to the position conservation information.
The results under refinement characteristic parameters
The ID algorithm was used to reduce the dimensionality of the amino acid component information, hydropathy component information, charge component information, secondary structure component information, and relative solvent accessibility component information to obtain a 10-dimensional ID value; the Position matrix scoring algorithm reduced the dimensionality of the position amino acid, position hydropathy, position charge, position secondary structure and position relative solvent accessibility to obtain a 10-dimensional S value. The obtained 10-dimensional ID value and 10-dimensional S value were collectively recognized as the 20-dimensional refinement characteristic by the SMO algorithm, and the results (OUR'S) by 5-fold cross validation were shown in Table 5.
At the same time, for the sake of comparison, the results of the SVM algorithm in paper [17] and the calculation results of SMO using the characteristic parameters of literature [17] were also included in Table 5.
As seen, the four acid radical ion ligands under the refinement characteristic parameters were very good, the MCC values were over 0.460, and the Acc values were all greater than 73.0%. Compared with the recognition results of all component information and all position conservation information, the values of S n , S p and Acc were gradually improved, indicating that the detailed characteristic parameters contain more complete information.
The MCC values of Zn 2+ , Fe 2+ , Fe 3+ and Cu 2+ have reached above 0.7, the MCC values of Mn 2+ and Co 2+ exceed 0.6, and the MCC value of K + was only 0.362; the MCC values of the eight metal ion ligands of Zn 2+ , Cu 2+ , Fe 2+ , Fe 3+ , Mn 2+ , Na + , K + and Co 2+ were improved in a small range compared with the results in Table 4, indicating that the eight ion ligands were more sensitive to the refinement characteristic; the evaluation indexes of Ca 2+ and Mg 2+ with the refinement characteristic parameters were not higher than that with the position conservation information, indicating that these two ion ligands were more sensitive to position conservation information; the Na + and K + have higher MCC values when the refinement characteristic was used as a parameter, compared with the results of all component information as characteristic parameters, it can be understood that Na + and K + were more sensitive to all component information under three characteristic parameters, but still lower than the results of other metal ion ligands, the MCC values of the residual ion ligands under the refinement characteristic parameters were improved compared with the results of all component information, which was the best results under the three characteristic parameters.
In general, the recognition result under the refined characteristic parameters was generally higher than the recognition result under the single combination characteristic parameter, which fully demonstrated that the compatibility performance of the SMO algorithm is good. In addition, new characteristic parameters were added based on the SMO results, and the prediction results for some ion ligands were improved, that is, the results of OUR'S in Table 5, indicating that the new characteristic parameters we added were useful parameters, suitable for the SMO algorithm.
Overall, in the process of ion ligand binding sites prediction, the SMO algorithm adopts analytical method to avoid the error accumulation caused by iteration method, so the accuracy of the prediction result is guaranteed; the PUK kernel function of this algorithm can deal with the nonlinear classification data of the binding sites prediction well and reflect the distribution characteristics of the training sample data, since it maps features from low-dimensional space to high-dimensional space, and achieves linear separability. Therefore, the SMO algorithm has a good performance for the prediction of ion ligands.
Conclusion
In this paper, the ligand binding sites of four acid radical ions and ten metal ions were predicted. Firstly, BioLip database was selected, and the optimal window sizes were determined by calculation; secondly, component information, position conservative information and detailed characteristics were extracted as characteristic parameters; then different characteristic parameters were input into the SMO algorithm, under the 5-fold cross validation, the identification of four kinds of acid radical ion ligand binding sites got a good result, among the results of the identification of ten metal ion ligands, the prediction results of transition metals were better than those of alkaline earth metals and alkali metals, the results of all position conservation information as characteristic parameters were better than the results of all component information as characteristic parameters, the prediction results under the refinement characteristic were better than the prediction results under the single combination characteristic, so the characteristic parameters can be refined as much as possible in the subsequent work.
Availability of data and materials
If you need data and materials, you can contact the corresponding author.
Ethics approval and consent to participate Not applicable.
Consent for publication
Not applicable. | 5,724.6 | 2019-12-01T00:00:00.000 | [
"Biology",
"Chemistry",
"Computer Science"
] |
A Real-Time Underwater Acoustic Telemetry Receiver With Edge Computing for Studying Fish Behavior and Environmental Sensing
Underwater acoustic telemetry has emerged as a powerful tool for practical applications, including resource exploration, environmental monitoring, and aquatic animal tracking. However, current acoustic telemetry systems lack the capability to transmit the collected data continuously in real time, primarily because the acoustic networking bandwidth is limited. Retrieval of the recorded measurements from the deployed receivers usually must be manual, leading to long delays in data retrieval and processing, high operational costs associated with the required manpower, and safety risks for the operators. In addition, there is no efficient way to continuously assess the status of the acoustic telemetry system, including the acoustic transmitters and receivers. Here, we describe the design, implementation, and field validation of a cloud-based, real-time, underwater acoustic telemetry system with edge computing for estimating fish behavior and monitoring environmental parameters. The system incorporates microcontrollers for edge computing and connects to a cloud-based service that further post-processes the transmitted data stream to derive behavior and survival information of tagged animals. The developed system has been demonstrated to have significantly improved performance over the benchmark system because of the integration of edge computing, with a greatly reduced energy consumption of 0.014 W resulting in the energy used by the acoustic modem being reduced by over 300 times. This work opens up new design opportunities for future real-time and multifunctional underwater acoustic systems.
I. INTRODUCTION
U NDERWATER acoustic telemetry finds diverse applications in fields of subsea resource extraction, such as resource exploration [1], tracking the movements of aquatic and marine animals [2], [3], wireless communications for autonomous underwater vehicles (AUVs) [4], and the emerging Internet of Underwater Things (IoUT) [5]- [8]. However, unlike terrestrial wireless communication systems, stateof-the-art acoustic telemetry systems suffer from limited data rates and energy efficiency, because of the characteristics of underwater communication, including bandwidthlimited underwater acoustic channels-generally recognized as the most challenging communication media in use today [9]-caused by high path loss [8], time-varying multipath propagation [10], large and variable propagation delay, node mobility caused by water currents [1], and Doppler spread associated with the complex underwater environment [11]. In addition to the fundamental constraint imposed by the characteristics of underwater sound propagation, there are system constraints that affect the operation of underwater acoustic communications. For instance, acoustic transducers have their own bandwidth limitations. For instance, the upper bit rate for the ATM-903 series (Teledyne Technologies, USA)-a widely used underwater acoustic modem-is 15 360 bits/s [12], far lower than the bit rate of terrestrial wireless communications. For many applications, the mode of operation associated with this bit rate is not suitable as it sacrifices most of the ability for the modem to detect and correct for communication errors and, as a result, significantly lower bit rates must often be used.
As a direct consequence of those limitations, end users cannot communicate interactively with the data source to manage the large streams of complex data; this inevitably leads to high latency and the high operational cost associated with the required manpower, as well as safety risk, for the operators manually recovering data. Furthermore, the harsh environment makes it more difficult to recharge or replace battery-powered acoustic telemetry receivers. Therefore, high energy efficiency is desirable for underwater acoustic telemetry systems [13]. For instance, one of the most important applications of acoustic telemetry in the Pacific Northwest region of the USA has been evaluating the behavior and survival of juvenile salmonids migrating through hydropower This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ dams in the Federal Columbia River Power System using the Juvenile Salmon Acoustic Telemetry System (JSATS) [14], [15]. The current implementation of the JSATS requires manual recovery of the data stored on secure digital (SD) cards from each of the deployed autonomous acoustic receivers at remote locations. If multiple JSATS autonomous receivers in an array fail, the detection efficiency of tagged animals could be significantly reduced [16]. In addition, once the autonomous receivers are deployed, there is no way to identify whether they are continuing to operate correctly.
The past decade has seen a tremendous research effort to improve underwater acoustic communication to achieve higher data rates and higher energy efficiency. One straightforward approach for higher data rates is to increase the signal bandwidth (or transmission/signaling rate) [17]. But an increase in signaling rate leads to larger intersymbol interference that degrades the quality of the received signal and requires compensation such as channel equalization. Instead, an efficient use of underwater acoustic communication bandwidth has been recognized as the pathway to improving its performance [17]. A software-defined underwater acoustic modem was proposed that has real-time adaptation capabilities to reconfigure its physical layer in real time under varying environmental conditions [18]. A bidirectional decision feedback equalizer was proposed, which exhibits improved signal-to-noise ratio (SNR), especially for timevarying channels [19]. Improved power efficiency was demonstrated using the orthogonal signal-division multiplexing technique over multiple transducers [20]. Novel data transfer schemes and networking protocols were proposed for higher energy efficiency and channel utilization [21]. Multiple-inputmultiple-output techniques were explored to increase the data rate over the bandwidth-limited channels [22]. However, the highest data rates that are offered by commercially available acoustic modems could achieve up to a few kilobits per second over distances up to a few kilometers, far below their terrestrial counterpart [5]. Furthermore, the proliferation of IoUT equipment to enable a new generation of underwater monitoring applications, with increasingly complex device structures and large amounts of data, has further intensified the computing resource demands and pushed the horizon of a new computing paradigm.
Edge computing is an emerging networking philosophy in which computational loading is shifted from the centralized cloud or end users to edge sensors where the data are initially collected [23]. Research and practice on this emerging technology have led to a wide spectrum of applications, including live video analytics, agriculture, and smart home and industrial Internet of Things (IoT) devices [24]. Recently, Yang et al. [3] have developed the Lab-on-a-Fish: a miniaturized biotelemetry tag that combines edge computing for studying the aquatic animals with improved efficacy. Edge computing offers a new strategy for shorter latency and effective bandwidth usage for underwater acoustic communication [25].
In this work, we developed a cloud-based, real-time underwater acoustic telemetry receiver with edge computing and connectivity to end data users. Edge computing enables the system to provide more effective data analytics with lower latency. The system incorporates microcontrollers and integrated circuits to create real-time data communication and processing units for edge computing; this is connected to a cloud-based service that further post-processes the transmitted data stream to derive the behavior and survival of tagged aquatic animals. The developed system has been successfully validated in a field environment, with a peer-to-peer acoustic communication range of up to 3.5 km and data communication to the cloud service in near real time. This work opens new design opportunities for future real-time and multifunctional underwater acoustic systems that can be used to estimate the behavior or survival of the tagged animals in near real time.
In summary, our major contributions include the following. 1) We identified the challenges in the existing acoustic telemetry solutions, namely, the contradiction between large volumes of raw data to be transmitted and the limited bandwidth of underwater acoustic communication. 2) We proposed the integration of edge computing to compress the large volume of raw data to a significantly reduced amount, which dramatically reduces the energy burden and increases the system's lifetime, enabling real-time communications from remotely located acoustic receivers. 3) We designed and implemented the proposed system and conducted laboratory and field experiments to demonstrate its feasibility and efficiency.
A. System Overview
We developed a cloud-based underwater acoustic telemetry receiver employing edge computing with increased energy efficiency and direct connectivity to end data users to enable real-time fish behavior and environmental monitoring. Fig. 1 is a system-level schematic diagram of the developed system, which consists of the offshore and nearshore/onshore (hereafter referred to as onshore) components.
The offshore components are responsible for receiving the acoustic signal from nearby fish-borne acoustic transmitters [26], performing edge computing, taking environmental measurements using off-the-shelf sensors, and transmitting the data to the onshore components. We designed a miniaturized, fully integrated printed circuit board (PCB). This PCB and the acoustic modem are incorporated into the existing housing of the offshore acoustic receiver. The PCB runs a custom-designed, real-time operating system that acquires the raw data from the acoustic receiver, removes common false positives using multipath and pulse repetition interval (PRI) filtering algorithms [27], obtains environmental sensor measurements [28], and compresses the resulting data into a transmittable packet.
The master onshore components handle five tasks. 1) Request and receive data from each acoustically linked offshore component. 2) Configure modem settings, such as transmission speed and power level. 3) Perform environmental measurements (temperature, barometric pressure, humidity, and wind direction). 4) Notify the slave onshore components to acquire the data from the offshore components they are linked to. 5) Send the data to the cloud service (Microsoft Azure Cloud Service) through either cellular or satellite communication, depending on service availability. The cloud-based system process visualizes the collected real-time data and provides options to view historical data. The embedded cloud services transmit the real-time data from receivers into the database and propagate it to the frontend.
We developed the front end of the cloud-based system using Angular 4, a framework from the Google Angular team and other communities, and the back end using JAVA EE. We chose GlassFish as the application server, and the database we used was PostgreSQL, an open-source relational database. To make the system platform independent, we created RESTful (representational state transfer) Web services for the communications and data transfers between the front-end user interfaces and the back-end database. For each receiver, we created a WebSocket to handle the realtime calculation of collected data for fish behavior. Google Maps was used to display the locations of the deployed systems.
B. System Implementation
The offshore hardware ( Fig. 6(a), see the Appendix) consists of: 1) a custom-designed PCB featuring a low-power, 32-bit microcontroller unit (MCU) as the core of the system; 2) an autonomous acoustic receiver; 3) a water leak sensor; and 4) an underwater acoustic telemetry modem for long-distance (up to 3 km in an ideal scenario) wireless communication in an aquatic environment. The autonomous acoustic receiver is equipped with a hydrophone that receives the high-frequency mechanical vibrations from tagged fish and converts them to electrical signals. The integrated digital signal processor (DSP) digitizes the input signal for use by the DSP in its detection and decoding algorithm. The detection algorithm searches for a tag signal, and the decoding algorithm determines what specific tag code is present. When a tag code is validated by the DSP, it passes the detection information on for storage in memory. The acoustic receiver is also equipped with built-in sensors for temperature, tilt angle, pressure (available but not included for this prototype), and battery voltage. The decoded tag detections and the data from the built-in sensors are transferred to the MCU in real time through an RS-232 serial interface. While developing this module, we also integrated a leak sensor, which sends a warning message to the cloud service, via the onshore system, when water is detected inside the acoustic receiver package. Since this module operates independently of the acoustic receiver, custom sensors can be readily integrated into the system without requiring modification of the hardware or software of the acoustic receiver.
The firmware was implemented as an interrupt-driven system for minimizing the power consumption of the offshore components. Fig. 7(a) (see the Appendix) shows a simplified flowchart of the firmware implementation. If no data are coming from the acoustic receiver at the input serial port of the microcontroller board, the board remains in the sleep mode, maintaining minimal functionalities to conserve power. When real-time data are received at the input serial port, the board wakes up, and the onboard timer is reset. The timer will accumulate if no data are seen at the input, and the microcontroller board will be put back to sleep after a programmed threshold. The firmware checks the incoming data and only starts to process the data if the input format fulfills the criteria. After all the data are processed, the microcontroller transmits the edge computed data to the offshore acoustic modem and resumes the sleep mode. The detailed description of the data processing scheme is as follows. First, the received data are parsed to extract the detected tag code (along with its detection time) and sensor data, including environmental temperature and the battery level of the acoustic receiver. Then, the extracted tag code is checked against a dynamic table of tag codes to search for the table index. If the extracted tag code does not exist in the table and the table is not full, the tag code will be filled into an empty slot of the table and its detection time is stored. Next, the detection time of the extracted tag code in the dynamic array is passed to the multipath filter. If it passes the multipath filter, it will be appended to an array and checked by the PRI filter. Otherwise, this data will be thrown away. For each tag code, it will skip the PRI filter if it has passed the PRI filter during the current hour since only hourly data for each tag code are sent. Further explanation and implementation details of the multipath and PRI filters will be discussed later. Finally, the firmware cleans up the table each hour by removing tag codes that have not been detected over the past hour and their corresponding detection times. The dynamic table ensures the firmware works with a large number of tag codes, with the only limitation being the maximum number of tag codes detected in a single hour. This maximum number of tag codes is determined by the size of the MCU's static random-access memory and can be improved by implementing external memory such as an extended ferroelectric random-access memory. Fig. 6(b) (see the Appendix) shows the hardware architecture for the onshore components. The core of the onshore components is a low-power microcontroller board running a real-time operating system. The microcontroller board is connected through serial ports to several components: 1) an underwater acoustic telemetry modem; 2) a cellular modem and an Iridium satellite modem for uploading the data to the cloud-based service; and 3) a weather station. The weather station integrates humidity and temperature, barometric pressure, and wind speed/direction sensors. The cellular modem can be replaced with a Wi-Fi modem or Ethernet modem for less-remote locations where these may be available. The cellular, Wi-Fi/Ethernet, or satellite modems pass the data packet through an intermediate service to our cloud-based service. The acoustic modem operates on a band between 22 and 27 kHz and has a source level of 178 dB. The modem is configured to run at a baud rate of 300 bits/s for enhanced reliability in a low signal-to-noise environment. Details of the components used for both onshore and offshore hardware implementation are shown in Table I. A finite-state-machine model was adopted for the firmware implementation of the onshore components, which is responsible for performing all control, monitoring, and wireless communication functions. Fig. 7(b) (see the Appendix) shows the simplified flowchart of the firmware implementation. First, the board is initialized with hardware configurations, including serial communication and an onboard real-time clock (RTC). The onboard RTC tracks the current time and wakes the system at the preprogrammed time (usually on an hourly basis). Then, the firmware checks whether the cellular module is awake or not. If not, the software tries to wake it up. Afterward, the board sends a command to the onshore acoustic modem to request data from each of the offshore acoustic modems. Data are requested from each offshore component one at a time, by sending the request to the modem address associated with the specific offshore component, and the onshore component will listen for a predefined duration for the incoming data. If no data are received within that preprogrammed time frame, the board will request data to the acoustic modem again. If no data could be received after a maximum number of retries, the board will stop sending data requests to the modem and proceed to the next step. After cycling through all the acoustically connected offshore components, the master onshore components notify the slave onshore components on the other side of the riverbank to acquire the data from the offshore components associated with it. This sequential data acquisition process ensures that no conflict would happen (i.e., the master and slave onshore components will not try to communicate with offshore components simultaneously). After collecting the environmental data from the weather station, the board attempts to send the received data to the cloud if valid data have been received or alternately send a warning message if a maximum number of request retries has been reached. The firmware will primarily try to send the data to the cloud through the cellular network. If the cellular network is not available, the data transmission will return failure and trigger satellite communication. This dual-communication mechanism ensures that our system is accessible in even the most remote locations. In most cases, cellular communication will be preferred because it is more cost effective. In places without cellular network coverage, or when the cellular modem fails, satellite communication can be used. Eventually, the board enters the sleep mode to conserve power. The program loops back to the beginning of the finite state machine and the same sequence is repeated at a preprogrammed time.
C. Edge Computing
The complex aquatic environment hampers underwater acoustic communication in several ways, including high path loss, time-varying multipath propagation, large and variable propagation delay, and Doppler spread. Given the low baud rate of acoustic modems (maximum speed of 15 360 bits/s with minimal error detection/correction), it was not feasible to transmit all the raw detection information, even after filtering out false detections. To address these issues, we implemented edge computing at the offshore acoustic receiver, including a sequence of filters to remove false-positive decodes that do not match the expected transmission interval (i.e., PRI filter) and further compression, as shown in Algorithm 1.
A schematic illustration of the filtering steps and their criteria is shown in Fig. 2(a). The original data set received by the acoustic receiver consists of a series of decodes at different times. Decodes are first processed using the multipath filter to remove decodes that originate from reflections of the acoustic transmissions (off the surface, bottom, or any other structures) or from transmission refraction in the water. The real-time implementation of the multipath filter is illustrated in the simplified flowchart shown in Fig. 2(b). Each time a tag message is received, it will be checked against the dynamic lookup table. The program then checks the number of messages for each valid tag code that was received. If the number of messages is equal to zero (either there was no message previously received, or the memory was cleared), the detection time of this message is kept as the time of arrival (TOA) of this tag code. Then, the time difference between this message and the most recently received message of the same tag code is calculated, and the time difference is compared with the multipath filter time window. In the current implementation, a value of 0.5 s (while the tags have a nominal PRI of 3 s) was used as the multipath filter time window. If the difference is smaller than the multipath filter time window, a multipath event is confirmed, and this message is ignored. Otherwise, we increment the number of detected messages for this tag code, and we save this detection time in an array.
We implemented a real-time PRI filter [ Fig. 2(c)] to filter out false-positive decodes for each tag code using the expected pattern of transmissions. The time difference of each message between a set of messages and the TOA of the initial message is calculated. This time difference is then sorted into one of three pairings of the number of messages (N m ) and the PRI window. The PRI window is based on a multiple of the tag's nominal PRI value and is defined as the nominal PRI × 1.3 × 12 + 1 s) [29]. If the time difference is smaller than the PRI window, and if N m is larger than the minimum number required to pass the PRI filter (N PRI ), i.e., N m ≥ N PRI , we proceed to the next step. If the time difference is larger than the PRI window, and if N m − 1 ≥ N PRI (the last message is dropped out since it is outside the PRI window), we proceed to the next step. If N m − 1 < N PRI , then the initial message is invalid and discarded and the filter moves on to the next message. After the PRI window condition check, the time difference between the initial message and each subsequent message is calculated, and an estimated PRI for each end if 6: process sensor data 7: extract the tagCode and detection time 8: if time difference< multiPathWindow then 9: calculate candidate PRI for each message 10: truePRI = mode(candidatePRIArray) 11: if truePRI != 0 then 12: count number of valid messages 13: if # of valid messages ≥ numPRI-1 then 14: passedPRIFilter = true 15: end if 16: end if 17: end if 18: end if 19: end while 20: compress data 21: prepare header, payload, and CRC code 22: send compressed data message is calculated. The mode of the estimated PRIs is computed to obtain the real PRI of these N m messages. Each estimated PRI is further compared with the real PRI, and only messages with a difference below a threshold (0.006 s in the current implementation) are considered valid. When at least N PRI valid messages from a set of messages match the PRI criteria within a predefined PRI time window, this set of messages passes the PRI filter. Otherwise, the initial message is discarded and the filter moves on to the next message. Note that the required N PRI (typically 4, 5, 6, or 7) depends on the acoustic environment. After the PRI filter is applied, and continuous messages are formed into a detection event. Each detection event contains at least N PRI messages and represents a continuous or nearly continuous sequence of transmissions.
D. Data Compression
We further reduce the data to a minimal data packet (Fig. 3), which is essential to accommodate the limited baud rate and to prolong the lifetime of a battery-limited underwater acoustic telemetry system [30]. Each data packet structure contains the header information, the payload, and the cyclic redundancy check (CRC). The data packet structure contains the following parts.
3) Temperature: 7-bit data within a range of 0 • C-25.4 • C to cover the temperature variation in the river over the year. 4) Voltage: 5-bit receiver battery voltage in a range of 1.9-3.5 V. 5) Total Number of Detections: 12-bit total number of raw detections (divided by 1000) that have been detected since the node was deployed, which serves as a replacement for the capacity used on the memory card. 6) Sensor: 16 bits, reserved for the potential environmental sensor. 7) n: 9 bits to represent the number of the tag code to transmit at this hour. 8) Data: 8 × ceil(n × 18/8)-bit data, which covers all the detected tag codes over the past two hours and their 2-bit tag history. n stands for the total number of tag codes. The first bit of the tag history indicates whether the tag was detected in the earlier of the past two hours, and the second bit indicates whether the tag was detected during the most recent hour. 9) CRC: 8-bit CRC code to handle error checking. An 8-bit CRC computation was implemented in the hardware (CRC-8-Dallas/Maxim, polynomial representation: . This implementation of a data packet does not require preknowledge of what tag codes are expected, which provides flexibility to handle a larger number of potential tag codes.
A. Laboratory Validation
The developed cloud-based underwater acoustic telemetry system was validated in a laboratory environment. Fig. 4 shows the results of tag detection, environmental sensing, and the remote receiver's health monitoring using the underwater acoustic telemetry system. The lab testing was conducted in an acoustic water tank filled with freshwater [29]. A receiving hydrophone from the acoustic receiver (SR3017, ATS, USA; electronics nearly identical electronics to ATS SR3001) was submerged in the tank 0.43-m deep and 1 m away from the transducer. The transducer was controlled by a workstation (Precision T7500 workstation, Dell, USA) via a MATLABbased interface to emit acoustic signals containing different JSATS transmitter tag codes [31]. The tag codes were sent with a 3-s PRI from the MATLAB interface via a data acquisition card (PCI-6111, National Instruments, USA). A total of 400 tag codes were transmitted, and each tag code was transmitted eight times in a row before the system switched to the next tag code. The offshore microcontroller board was connected to the acoustic receiver to obtain the raw detection data. The microcontroller processes this data, forming an hourly summary of tag, receiver, and sensor data and then transmits this summarized data to the onshore acoustic modem serially connected on the other end. Fig. 4(b) shows the measured temperature of the water tank during 72 h. A stable temperature of around 20 • C was obtained since the lab is a temperature-controlled environment. The unwavering battery voltage of around 3.5 V, as shown in Fig. 4(c), was a result of the acoustic receiver being connected to a lab power supply. About 63 000 detections were captured for the 400 valid tag codes, which complies with the input MATLAB program. Using the developed system, we visualized the detected tag codes in real time. These are shown in Fig. 4(e), where a green box stands for a tag code that was detected and passed the multipath and PRI filters.
B. Field Validation
The developed system was subsequently validated in a field environment. The first field trial took place on August 26, 2020, in Levey Park, WA, USA [ Fig. 5(a)]. This field environment was chosen as it is close to the Ice Harbor Hydroelectric Dam. The Ice Harbor Dam is located on the Snake River, approximately 14 km from the mouth of the Columbia River, where there is particular interest in understanding the impacts dams have on the survival and behavior of migrating juvenile salmonids [32] protected by the Endangered Species Act of 1973.
Both the offshore and onshore components were deployed off the docks at a depth of 1.5 m and approximately 50 m apart. A custom acoustic transmitter, which transmits a list of tag codes to the offshore component at a PRI of 3 s, was deployed off the same dock as the onshore component. The total duration of the data collection and edge computing was about 2 h. The tag code list and sensing data were successfully communicated to the Azure server [ Fig. 5(b)], showing that the acoustic telemetry system functioned as expected. Thanks to the developed onboard edge computing algorithms, a total of 458 000 bytes of raw data (which contained 36 valid tag codes after the PRI filter) was compressed to only 200 bytes. To estimate the deployment time of the acoustic modem, the following equations are used: where the constant 24 is derived from 24 h/day, converting hours to days, and the constant 450 is derived from 8 bits/byte and 3600 s/h, converting bits/sec to bytes/hour. Detailed information of the parameters used for the calculations is shown in Table II.
Because of the integration of the edge intelligence, our developed system shows a significant performance improvement over the benchmark system (ATM-903 series, Teledyne Technologies, USA), with a dramatically reduced energy consumption of 0.014 W (in comparison to 4.2478 W), and an estimated lifetime increase by over 300 times (892 days in Table III. Here, the benchmark system refers to what we would expect from the acoustic modem being connected directly to the full raw output of the acoustic receiver, i.e., no edge computing to reduce the amount of data to be transmitted, but the modem operating in an identical configuration. To confirm the correctness and accuracy of the edge computing, we ran the MATLAB script from the previously designed offline data processing algorithm on the data recovered from the acoustic receiver and compared the result with the results from edge computing. The edge computing showed a perfect agreement with the offline data processing algorithm. The quality of the acoustic link between the onshore and offshore components was field tested by instructing the onshore component to transmit one preset test message at a specified transmission rate (Tx rate) and power. The resultant transmission time as a function of the Tx rate, automatic gain control (AGC), SNR, and multipath delay (MPD) as a function of transmission power (Tx Power) are shown in Fig. 5(d)-(f), respectively. By using a communication speed of 1000 bits per second (bps), a data packet of 1000 bytes (which corresponds to 180 detected tag codes) could be transmitted within 11 s. The MPD shows an increasing trend as the Tx Power increases. The received signals need less AGC as Tx Power is strengthened. Surprisingly, a higher Tx Power has caused a decreased SNR.
To investigate the maximum communication range of the developed system, we carried out a second field test on September 23, 2020, at the same location [ Fig. 5(g)]. The offshore component was fixed at a single location throughout the experiment, while the onshore component was temporarily deployed from a boat at each test distance. The offshore component was anchored at the bottom of the river at a depth of about 25 m (location 1 in Fig. 4(g)). The onshore component was lowered into the river from the boat at a depth of about 5 m at locations 2-8 in Fig. 5(g). Eight communication ranges were tested: 1) 0.5 km; 2) 1 km; 3) 1.5 km; 4) 1.5 km; 5) 2 km; 6) 2.5 km; 7) 3 km; and 8) 3.5 km. The system functioned successfully at all eight tested locations, although the communication became observably unreliable at the 3.5 km distance. The SNR shows a clear dependence on the communication range, as displayed in Fig. 5(h). At the communication range of 3.5 km, an average SNR of around zero was obtained.
IV. CONCLUSION
In this study, a real-time acoustic telemetry receiver system with edge computing for studying fish behavior and environmental sensing was designed, implemented, and validated in a controlled field environment. Edge computing was implemented at the acoustic receiver to allow the data to be preprocessed and filtered closer to where it is created. The system integrated a flexible connection to the cloud via cellular, Wi-Fi, or satellite communication, to allow the most cost-effective solution that still allows the system to be deployed in remote locations. These calculated metrics could be used to elucidate the overall status of fish migration during a given season and allow hydropower operators to observe the effects of operational changes, such as changes to spill patterns, in near real time. The system was implemented with the flexibility to allow additional sensors to be integrated into both the onshore and offshore components for environmental monitoring. The system has been tested and validated in both a laboratory and a controlled field setting. The maximum operational range of the system was confirmed to be 3.5 km in a field environment. In addition to allowing near-real-time detection information, this system will allow remote health monitoring of each of the receivers deployed. This will allow users of the system to detect issues that may significantly reduce the performance of the system. The system described in this contribution is expected to bridge diverse technologies for sensing the ocean and contributing to the development of IoUT and Smart Ocean. Possible future extensions and applications of the system include: 1) a system that combines the onshore and offshore systems for applications where we want the realtime data but can deploy from the surface and 2) a system that implements a customized version of the receiver firmware that outputs slightly less information but close to true real time for applications such as using the system on an autonomous robot. integrate an acoustic modem and assistance in interfacing the custom electronics designed in this study with their hardware. | 7,620.2 | 2022-09-15T00:00:00.000 | [
"Computer Science"
] |
Axiomatic Design Based Complexity Measures to Assess Product and Process Structures
Definitions of complexity often depend on several circumstances, such as the nature of investigated complex system, the kind of complexity, the conceptual framework used for a study, the theoretical approach taken, and the like. In this paper, two complexity measures that are based on Boltzmann’s entropy concept and AD theory are proposed and described. The first measure quantifies product variety complexity and the second one focuses on process structure complexity. Such complexity techniques will be used to determine product platform complexity and related process complexity for early stage of design decision-making. The method focused on product platform complexity assumes that the distribution of FR-DP couplings offers a suitable complexity concept, which prescribes that coupled designs should be decoupled, if possible, since uncoupled design is ideal and a decoupled design is less good, while a coupled design is the least satisfactory. Analogically, the same principle is used for the purpose to quantify topological process complexity by transforming input components into process variables and product modules including final product into design parameters. Subsequently, relevant properties of these measures will be analysed by computational experiments. Finally, practical findings for mass customization practice will be presented.
Introduction
Definitions of complexity often depend on several circumstances, such as the nature of investigated complex system, the kind of complexity, the conceptual framework used for a study, the theoretical approach taken, and the like.Overviews of complexity approaches and theories were offered by number of studies, see, e.g., [1][2][3][4] which provide complementary information among them.According to Gao et al [5] Shannon information concept seems widely recognized as essential building block of complexity theory.Traditional complexity metrics are associated with an absolute measure of complexity in contrast to the axiomatic design complexity, where this system attribute is treated as relative quantity based on the information concept and axioms of Axiomatic Design (AD).The complexity is expressed as: "A measure of uncertainty in understanding what it is we want to know or in achieving a functional requirement (FR) [6]."A relative information quantity in this approach is determined by the overlap between the system range of FRs and the design range of FRs.By this complexity theory, four different types of complexities: time-independent real complexity, time independent imaginary complexity, time-dependent combinatorial complexity, and time dependent periodic complexity are identified.The key idea of the theory is reduction of a complexity of system design in order to increase system reliability at each level of the design hierarchy.The dependencies between the FRs and the design parameters (DPs) can be classified by three types of design matrix (DM): uncoupled, decoupled and coupled.Among them, coupled design where FRs are influenced by possible changes of individual FRs, is more complex than an uncoupled design.Guenov [7] extends AD complexity theory for computing the information content of all types of design matrices.For this purpose, he developed complexity measures of design representations by adopting Boltzmann's entropy concept.Existing research literature on AD (see, e.g.[8][9][10][11]) offers other challenging approaches and inspirational studies including complexity issues.In this paper, two complexity measures are proposed.The first one, determines product variety complexity and the second one is dedicated to measure process structure complexity.Both of them adopt one of the complexity measures from the work by Guenov [7].Such complexity measures will be used to define product platform complexity and related process complexity for early stage of design decision-making based on the assessment of topological structures of design matrices.The method focused on product platform complexity assumes that the distribution of FR-DP couplings offers a suitable complexity concept, which prescribes that coupled designs should be decoupled if possible, since uncoupled design is ideal and a decoupled design is less good, while a coupled design is the least satisfactory [6].Analogically, the same principle is used for the purpose of quantifying static process complexity based by transforming input components into process variables and product modules including final product into design parameters.
Theoretical background
Adapted complexity measures used in our approach are based on Boltzmann's formula [12] for statistical entropy introduced in 1872.This formula in case of an ideal gas exactly corresponds to the thermodynamic entropy.According to Anderson [13] Boltzmann's entropy is simply related to disorder in many field of science, even though this resemblance for the most part have nothing to do with the second law of thermodynamics.Guenov [7] substituted number of elements (molecules) in formula for degree of disorder (Ω) by total number of couplings (N) in design matrix in order to calculate design complexity based on the assessment of topological structure of the design matrix as follows: where K is number of design parameters (number of columns in design matrix).
The same author recognized that formula: which was derived from the statistical entropy seems to convey better the meaning of axiom 1 of AD, where Nj is the number of couplings per design parameter (i.e. per column), j = 1,…, K. Therefore, this complexity measure denoted by us as Systems Design Complexity (SDC), will be applied for the purpose to estimate product variety complexity and process structure complexity.
A simple example demonstrating the application of this measure can be shown by using random design matrix (DM) with its couplings distribution depicted in Figure 1. 3 Adaptation of SDC to measurement of product variety complexity
Description of the product complexity measure
In order to adopt SDC measure for quantifying product variety complexity, the following steps are involved: 1) Classification of input components (ICs) entering an assembly process in terms of mass customization, 2) Description of graphical model interpreting relation between ICs and possible product (module) configurations (PPCs), 3) Conception of transformation mechanism of the relation between ICs and PPCs into a DM.
Classification of input components
Due to the fact the structure of variable product components determines total number of product combinations, it is reasonable to introduce working classification of input components.We consider three types of initial components entering an assembly process.They are as follows: Stable components (S) are considered to be assembled for ensuring the functionality of the module or final product.
Optional components (O) are useful in some cases but they are not required.They can be selected according to the customer´s requirements and are optional in any combination, including cases when only individual components are chosen.Selection by customer without this type of component is also an option.
Compulsory optional components (CO) are different from O by the number of components that may be chosen from all of them.They are limited in selection.Thus, restrictions are determined by three specific selection rules: minimum, maximum and exact requirements on selection.These selection rules can be specified in a simple way by combinatorial number , where l defines ways of picking component combinations from a set of all k, while 1 ≤ l < k.
In mass customization environment, practically any number of stable, voluntary and compulsory optional components can be combined.However, the following specific selection rules of the selections for the set of CO components with number k may occur when identifying product configurations.They are: Individual selectivity rule, where it is possible to define exact number of 'l' of components to be chosen from all 'k' of CO components; Maximum selectivity rule where it is possible to define the maximum number 'l' of CO components to combine within an assembly choice of all 'k' of CO components (note that 'l' is max.k-1); Minimum selectivity rule where it is possible to choose/combine at least 'l' CO of components from available 'k' of CO components (note that 'l' is min.1).
Combinatorial formulas for determining numbers of product (module) configurations according to the above described components types and selection rules are presented by Modrak [14].The example of customizable chair containing of three types of initial components is shown in Fig. 2.
Graphical interpretation of relation between ICs and PPCs
Subsequently, it is useful to model the relation between ICs and PPCs by incomplete bipartite graphs (see an example in Figure 3a) with two independent sets of vertices U and V, where the set U consists of ICs and the set V contains from related PPCs.The set U in given case involves three stable components (i=3), one optional component (j=1), and three compulsory optional components with the individual selectivity rule .Frequently, practitioners used to express product variety complexity by number of PPCs.However, the graphical models can be helpful to study product variety complexity in more detail way.
Transformation mechanism of scheme of relation between ICs and PPCs into a DM
The transit mechanism, which was partly outlined in our previous work [15], is based on substitution of set U consisting of ICs by DPs.Subsequently, elements of set V, i.e.PPCs will be replaced by FRs.Moreover, the transformation takes under consideration the fact that number of stable components does not impact on number of PPCs.Therefore, all stable components occurring as initial assembly inputs are represented only by one DP as shown in Figure 3b.Finally, the bipartite graph represented by FR-DP relations can be easily transformed into a system design matrix shown in Figure 3c.Then, product variety complexity can be enumerated by using formula (2) as follows: SDC= ΣNj ln Nj = 6ln6 + 3ln3 + 2ln2 + 2ln2 + 2ln2 = 18,21 nats.
This transformation assumes that DPs considered as inputs components are determined by individual customer's specific needs through FRs of the product.Such defined FRs are directly included in selected product configuration.Then, the model of FR-DP relations depicted in Figure 3b is coherent with three of four domains of the design world [16] as shown in Figure 4.
Comparison of SDC against PPCs
In order to analyse possible differences between two possible product variety complexity measures, i.e., SDC and PPCs, two scenarios will be considered.
Scenario #1
In this scenario, structure of ICs is assumed as combination of two and more S components and one and more O components.Let's compare SDC and PPCs by using numbers of these ICs as described in Table 1.As it can be seen from Table 1, number of S does not impact on PPCs and SDC values.This fact can be simply explained by transforming selected three combinations of ICs as shown in Figure 5.
Examples of combinations of ICs with different number of S.
In Figure 5, all stable components in original bipartite graphs are replaced only by one stable component.Then all three modified graphs are identical, which logically proves that SDC measure reflects this reality adequately.To find the answer, the example from Figure 6 will be used.By comparing two bipartite graphs for the rules and , it is empirically evident that product variety complexity for the rule is higher than for the rule .However, this practical view can be proved by formal verification.For this purpose, the following definition for structural complexity will be employed.The complexity of a system scales with the number of its elements, the number of interactions between them, the complexities of the elements, and the complexities of the interactions (Gershenson 2002) [17].By applying the previous definition for the two subscenarios #2.1 and #2.4 from the Figure 6, the following results can be obtained.( 5 1 ) ( 52 ) ( 53 ) ( 5 4 ) ( 61 ) ( 62 ) ( 63 ) ( 64 ) ( 65 ) ( 71 ) ( 72 ) ( 73 ) ( 74 ) ( 75 ) ( -number of nodes: 10, -number of edges: 30, -the complexity of the edges: one edge set -the complexity of the nodes: four component set Then, according to the previous complexity definition, it is proved that product variety complexity for the rule is higher than for the rule .
By analogy it can be proved that product variety complexity for the rule is higher than for the rule .
The same logic can be used to verify that SDC values better reflect product variety complexity than number of PPCs for different amount of compulsory optional components.
Adaptation of SDC to the process complexity measure
Similarly, as in the case of adaptation of SDC to measurement of product variety complexity, manufacturing process structure will be transformed into coupled design matrix.This transformation in subsequent subsection will be described.
Description of the process complexity measure
Prior to transformation of manufacturing process structure into DM, model of assembly process structure will be introduced.
Description of model of assembly process structure
Assembly type of operations are commonly interpreted using graph theory as convergent graphical models.Under convergent assembly structure we understand the chain where one process node has at most one successor, but has to have at least two predecessors.Our framework of assembly structures follows the work of Hu et al. [18], who outlined the way to model possible supply chain structures based on the number of original suppliers.An example of the model of assembly process structure is shown in Figure 8 a), where i=1,2,…,m is the number of stable input assembly components; O=1,2,…,p is the number of assembly operations, while O0 represents final assembly operation.
Transformation mechanism of manufacturing process structures into a DM
In this transformation, equally as in Subsection 3.1.3,stable input components will be substituted by DPs and process nodes will be replaced by PVs (see Figure 8 b)).Subsequently, the convergent graph represented by DP-PV relations can be easily transformed into a system design matrix shown in Figure 8 c).Then, process complexity can be enumerated by using formula (2) as follows: SDC= Σ Nj ln Nj = 3ln3 + 3ln3 + 2ln2 + 2ln2 + 1ln1 = 9,36 nats.
In such graphs, input assembly components are mapped from physical domain to process domain as can be seen in Figure 9.
Comparison of SDC against to concurrent measure
Competitive properties of the SDC measure can be found by its comparison with possible concurrent measures.
Possible alternative complexity indicators were already mutually compared in previous studies [19,20], where so called Index of vertex degree (Ivd) met optimality criteria for assessment of network complexity in the best way.Index of vertex degree has been introduced by Bonchev and Buck [21] and is expressed for Graph G consists of a set of V vertices, {V}≡{v1, v2, … , vV}, by formula: where deg(v) is the degree of vertex v in G.
Differences between these two process complexity measures, i.e., SDC and Ivd, will be analysed and evaluated through the following two computational experiments.
Description of computational experiment #1
The aim of the first experiment will be an investigation of complexity differences between assembly process structures resulting from product variants.For this purpose, real industrial case for assembly of chairs will be used.Customization of the chair includes colour and material variants of the seat panel, two modifications of the legs design, and three types of back support.Totally, there are 747 possible product variants.When material and colour varieties are omitted, there are three variants of product design and the same number of assembly process structures (see Figure 11).
Description of computational experiment #2
The aim of the second experiment is to show differences in sensitivity between the two indicators.The experiment is based on the real assumption that each of the 3 assembly structures can be topologically modified by splitting or integrating process operations.For example, when number of initial assembly components is 6, then number of all possible process alternatives is 33.When we compared complexity of all possible theoretical process structures for 4, 5 and 6 initial input components we found that there are the same complexity values for different process structures by using Ivd indicator.For such process structure, we applied SDC indicator as shown in Figure 12.As can be seen from obtained results in Figure 12, indicator SDC identifies different complexities between the pairs of the structures.It proves that SDC reflects the differences between these structures more sensitively.
The evaluations of the obtained results from computational experiments applied in the Section 3 and 4, indicate strong theoretical and promising practical potential of the two complexity techniques for the control and reduction of complexity in mass customization environment.The inherent properties of these two techniques for measuring observed complexity seem to be more suitable for given purpose than analysed concurrent indicators Ivd and PPCs.Moreover, both indicators fit into the theoretical construct of axiomatic design not only from the viewpoint of partial dependencies between FRs-DPs and DPs-PVs, but there are clear mutual relations between all four domains of the design world as it is shown by Figures 4 and 9.
Finally, we would like to point out that our both proposed techniques for measuring observed product complexity and process complexity at the same time validated systems design complexity metric by Guenov [7] using fundamentals of Architectural Design and Axiomatic Design for comparison of alternatives.Even though this complexity metric does not include the information axiom of AD, it disposes of useful practical applications.
Fig. 2 .
Fig. 2. Example with three types of initial components.
5 FR 1 FR 2 FR 3 FR 4 FRFig. 3 .
Fig. 3. a) Two independent sets of vertices U and V; b) Relation between DPs and FRs; c) Final transformation of the relation between ICs and PPCs into DM.
6 Fig. 4 .
Fig. 4. The relation between ICs and PPCs in context of four domains of the design world.
3. 2 . 2 .
Scenario #2 This scenario supposes several CO components with all possible individual selectivity rules.Let's use the example with five CO components and all possible individual selectivity rules, namely , , and as shown in Figure 6.
4 Fig. 6 .
Fig. 6.An example of scenario with 5 CO components with all possible individual selectivity rules.
Fig. 8 .
Fig. 8. a) Initial model of assembly process structure; b) Converted model of assembly process structure; c) Transformation of the converted process structure into DM.
2 Fig. 9 .
Fig. 9.The relation between input assembly components and process nodes in context of four domains of the design world.
Fig. 11 .
Fig. 11.a) Design variants of customized chair; b) Related process models of assembly operations with DMs.For example, assembly process structure No. 1 consists of 2 operations.The first one includes assembly of back support with compact legs, and subsequently seat panel is added.When applying two concurrent complexity measures, Ivd and SDC, for determination of topological complexity of the structures in Figure11 b) one can see that complexity values of both indicators have the same tendency and provide similar results.It justified the applicability of the measurement method of SDC for intended purpose.
Fig. 12 .
Fig. 12.An examples of process structures containing four, five and six assembly ICs with SDC and Ivd values.
Table 1 .
PPCs and SDC values for selected combinations of S and O components. | 4,323.2 | 2018-01-01T00:00:00.000 | [
"Engineering"
] |
Channel-Aware QUIC Control for Enhanced CAM Communications in C-V2X Deployments Over Aerial Base Stations
The proliferation of vehicular ad hoc networks necessitates efficient data transfer protocols, particularly in the context of Cellular Vehicle-to-Everything (C-V2X) communications. This paper focuses on enhancing the performance of the Quick UDP Internet Connections (QUIC) protocol, focusing on cooperative vehicular networks supported by aerial drone relays. While QUIC outperforms traditional protocols, its default congestion and flow control mechanisms do not adequately address the unique challenges posed by volatile networks spanning the terrestrial and aerial domains, as they are characterized by frequent topology changes, and high propagation delay volatility. We analyze QUIC's congestion and flow control and propose enhancements to optimize its performance in such networks, specifically designed for C-V2X communications in Open Radio Access Networks (O-RAN). Our proposal adjusts connections' congestion window size and individual streams' flow control windows in a channel-aware manner. Simulation experiments assess the performance of our proposal, comparing it with QUIC's default mechanisms. Our proposal can be seamlessly integrated into existing implementations, making it a viable approach for improving performance and addressing the challenges specific to vehicle-to-drone communications. By addressing QUIC's limitations and optimizing its performance for C-V2X applications in O-RAN, our enhancement offers a valuable contribution towards enabling low-latency, and resource-aware vehicular communications for the realization of autonomous driving and advanced vehicular services.
dynamic and transient network layer, new requirements and transmission provisioning necessities have emerged.In the current state of the art, we can categorize relevant protocols in two main groups, namely the reliable protocols, which offer trustworthiness at the expense of responsiveness e.g., TCP and SCTP, and the unreliable datagram protocols capable of offering great responsiveness at the cost of packet reception dependability, such as UDP, DCCP, and RTP.As a means of keeping the best of both worlds, the QUIC protocol was finalized by J. Iyengar and M. Thomson [1] in May 2021 and aims to improve upon traditional transport layer protocols.
While TCP is a reliable protocol that ensures data is delivered without errors, it suffers from high latency due to its threeway handshake and slow start algorithm.UDP, on the other hand, is a much faster protocol that does not guarantee reliable delivery but is often used for real-time applications such as video streaming and gaming.QUIC was developed with several pivotal features, such as multiplexing, encryption, and importantly, a flow control mechanism.The reasoning behind resorting to QUIC (which is a connection-oriented protocol) instead of adopting a connectionless approach) or adopting TCP directly can be further elaborated.For cooperative awareness messages (CAMs), QUIC, is able to overcome the persistent constraints of TCP's congestion and flow control mechanisms (e.g., slow start and head-of-line blocking) through stream multiplexing and perstream flow control which results in the preservation of high throughput and low latency in environments that are characterized by diverse topologies and dynamic network characteristics.Contrary to TCP, QUIC manages congestion windows and flow control parameters for each data stream independently, therefore making the data transmission more agile and flexible in comparison to TCP.Additionally, QUIC integrates TLS 1.3 directly into the transport layer, dramatically improving the security and leading to reduction of connection and transport latency-the most exceptional advantage in C-V2X communication where state messages are immediately exchanged within CAMs.The motivation to utilize QUIC for CAMs lies not only in its performance superiority against TCP, but also in comparison to connectionless protocols (with UDP being its transportlayer foundation).Choosing QUIC, in this case, can be seen as optimal for CAM delivery under the condition of complex and dynamic environments of vehicular communication.This connection-oriented strategy is the key point of it that offers more efficient method of communication management in high density networks where vehicles experience frequent link loss and links are either closed or reconnected.As a result, the network performance gets significantly degraded.QUIC's internal architecture is innately prepared to troubleshoot issues such as packet loss, changing bandwidth and spikes in latency, swiftly, thereby ensuring the timely and reliable delivery of the security-related data.What is critical is the presence of these features in order to be appropriate for services that depend on the ultra-reliable low-latency communications (URLLC), which is why QUIC is a very important element in vehicular networks where the traffic is always continuous, coherent, and synchronized to ensure real time.
In order to support the wider adoption of QUIC for mobile ad hoc environments, we aim to revisit the flow and congestion control mechanisms.Currently, QUIC's flow and congestion control mechanisms are sub-optimal for transient networks, which are characterized by frequent topology changes, unpredictable bandwidth, and high packet loss rates.We aim to propose an enhanced Flow Control function to implement stream multiplexing in a more efficient manner, considering the global connection optimum.Lastly, to improve congestion control, we propose using a modified version of the Swarm-HTCP (S-HTCP) additive increase-multiplicative decrease (AIMD) scaling algorithm for increased performance in high-mobility ad hoc network [2].Our approach creates a solid abstraction layer between the volatile network medium and the QUIC overlay, effectively highlighting the protocol's deterministic behavior.By isolating the protocol from the volatile network medium, we are able to ensure that the protocol's behavior is consistent and deterministic, which is essential for reliable and efficient data transfer.Overall, our approach provides a robust and reliable solution for enhancing the congestion and flow control functions in QUIC for volatile, vehicular networks.
We emphasize the significance of CAMs in the context of O-RAN and its potential impact on the simulation and modeling of Open Cloud and C-V2X networks.Our research aims to enhance the existing frameworks by integrating improvements that address key aspects such as the design of test-bed architectures and the simulation of Open Cloud and C-V2X scenarios.In our work, we utilize amongst others, the O-RAN E2 NS-3 module [3] facilitating support for running multiple terminations of an O-RAN-compliant E2 interface within a simulation.This integration allows us to explore the dynamics of O-RAN networks more comprehensively and assess the performance of CAM message exchanges in a realistic environment.By leveraging this capability, we can evaluate the effectiveness of O-RAN deployments, identify potential bottlenecks, and optimize the network configurations.Furthermore, our research incorporates the C-V2X mode 4 ns-3 module [4] specifically tailored for C-V2X Mode 4 communications.This model builds upon the ns-3 (Device-to-Device) D2D model from NIST, providing a reliable foundation for simulating and evaluating C-V2X scenarios within the O-RAN context.By employing this model, we can assess the efficiency of CAM message delivery, analyze the impact of network dynamics on communication reliability, and explore potential optimizations to enhance the overall performance of C-V2X networks.
Our proposed improvements enable us to push the boundaries of Open Cloud and C-V2X simulation and modeling frameworks.Accurate representations of O-RAN architectures (including support for O-RAN-compliant E2 interfaces and C-V2X Mode 4 communication) allow us to simulate realistic network conditions, evaluate performance metrics, and gain insights into the behavior and interactions of various network components.We utilize said modelling capacity to propose a C-V2X application-specific variant of the QUIC protocol, which is in turn evaluated using the same emulation framework.Ultimately, our research aims not only to propose a new variant of QUIC, but also to contribute to the advancement of O-RAN technologies, the development of efficient testbed architectures, and the simulation and modeling of Open Cloud and C-V2X networks.By bridging the gap between novel application-specific protocols and vehicular communications modelling solutions, we can facilitate the optimization of network designs and pave the way for the realization of reliable and high-performing vehicular communication systems.
The remainder of this work is structured as follows.Section II provides an outline regarding the reasoning and motivation for conducting this research while it also clarifies the contributions of our work to the underlying modules it utilizes.Section III provides a comprehensive description of QUIC's features and key mechanisms placing particular emphasis on the protocol's congestion and flow control mechanisms.Section IV describes the followed methodology for implementing our proposed enhancement, providing mathematical modelling for all involved components and elaborating on our proposed bi-fold improvement.Section V provides detailed information regarding the evaluation of the proposed scheme, including obtained results and a comparative analysis.Lastly, Section VI concludes our work after a comprehensive discussion and summary of the improvements achieved.
II. RATIONALE AND MOTIVATION
The rapid evolution of vehicular networks and their increasing role in enabling intelligent transportation systems necessitates the development of communication protocols that can adapt to the dynamic and often unpredictable nature of these environments.The QUIC protocol, originally designed to optimize web traffic efficiency through reduced connection and transport latency, presents a compelling foundation for vehicular communication due to its inherent advantages over traditional protocols like TCP.However, the direct application of QUIC in vehicular networks without considering the unique challenges posed by these environments could lead to reduced performance.Research in optimization methods specifically targeting the QUIC protocol's mechanisms in order to finetune them for usage in vehicular and/or beyond-5G scenarios have already showcased promising results [5] [6]; our work draws motivation from this research and takes it one step further through the introduction of channel awareness and aerial link-enabled relaying.We additionally consider the outputs of authors in [7] who also considered potential enhancements in QUIC's flow control mechanisms to achieve The inclusion of Unmanned Aerial Vehicles (UAVs) in vehicular communication scenarios is highlighted by recent research, reflecting their growing importance in the wireless network landscape.UAVs serve as aerial platforms, enhancing network resilience, extending communication ranges, and facilitating rapid network deployment [8], particularly in scenarios where ground infrastructure is lacking or compromised.Their utility is especially pronounced in emergency situations where traditional communication networks may be impaired by natural disasters or catastrophic events, allowing for swift establishment of ad-hoc networks to ensure uninterrupted communication for rescue and coordination efforts [9].Further advancing the integration of UAVs, recent studies have demonstrated their role in C-V2X communication between Connected and Autonomous Vehicles (CAVs) and UAVs [10]; utilizing a combination of communication technologies including Dedicated Short Range Communication (DSRC), User Datagram Protocol (UDP), internet-based WebSocket, and Transmission Control Protocol (TCP), aerial links can support various use cases, including accident location information sharing and real-time photo transmission from UAV-mounted cameras to traffic management systems.Figure 1 showcases the envisioned system model, showing the involved entities and an example of the data flow process for exchanging a CAM over ad hoc aerial links.
Such methods are expected to significantly enhance network density, support ultra-reliable low-latency communications, and enable advanced services, such as smart city applications, autonomous driving, and remote sensing [11].Our research aligns with these forward-looking perspectives, employing UAVs not just as theoretical elements but as practical enablers of robust, adaptable, and far-reaching vehicular communication systems.The primary motivation behind this study stems from the observation that the default congestion and flow control mechanisms of QUIC are not tailored for the high mobility and variable network conditions characteristic of vehicular networks, specifically considering aerial nodeenabled vehicular communication relaying, which is both a next-generation use case involving 5G networks where drones will play a pivotal role in ensuring constant overage of connected (vehicular) nodes [12] [13] [14], and a disasterrelief use case where already aerial nodes have seen relaying scenarios.These environments are marked by rapid changes in network topology, variable signal quality, and fluctuating network densities, all of which can severely impact the reliability and efficiency of communication protocols.We aim to enhance the QUIC protocol by integrating adaptive mechanisms that leverage real-time network conditions to dynamically adjust its operational parameters.This approach seeks to maintain optimal flow control and congestion management, thereby ensuring robust and efficient communication even in the face of volatility.
Our QUIC enhancements significantly advance vehicular communications, catering to the stringent demands of 5G networks and URLLC requirements [15].Tailored for vehicular network dynamics, these modifications ensure robust, lowlatency communications essential for autonomous driving and intelligent traffic systems.By integrating advanced flow and congestion control mechanisms, our protocol optimizes data transmission in highly mobile environments, crucial for effective emergency response and vehicular IoT applications.This not only demonstrates the protocol's real-world applicability but also its pivotal role in evolving vehicular networks towards enhanced efficiency and reliability.a) Our Contribution: Considering the motivation described above, we introduce significant advancements to the congestion and flow control mechanisms within the QUIC protocol, tailored to the unique and dynamic challenges of vehicular networks.Our contributions extend into the development and implementation of a novel congestion control algorithm, based on a modified version of our previous work (S-HTCP [2], and a pioneering flow control method, "Globaloptimum Aware Dynamic Flow Control", which is a unique addition proposed by our present research. 1) A novel Congestion Control Algorithm: The S-HTCP algorithm which we have further enhanced, represents a paradigm shift in the approach to congestion control within QUIC.This algorithm employs an AIMD strategy which considers the time elapsed after the last observed congestion event and the minimum Round-Trip Time (RTT) of a given flow.It has now been further developed to adjust its behavior based on a nuanced analysis of RSSI volatility.This allows for a more granular and adaptive response to network conditions.Our approach significantly improves upon traditional methods, providing a robust solution that addresses the rapid changes inherent in vehicular networks.This new congestion control mechanism is meant to replace the default algorithms supported by QUIC.2) Global-optimum Aware Dynamic Flow Control: Beyond congestion control, we also introduce a novel flow control mechanism designed to maximize the efficiency and fairness of resource allocation among multiple streams within a QUIC connection.This method addresses the common issue of individual streams under-utilizing their allocated flow control and congestion windows due to diminished Acknowledgement (ACK) rates, which can lead to bottlenecks.By considering the global optimum of at a connection level, our approach mitigates these bottlenecks and ensures a more efficient and equitable distribution of resources.This innovation enhances the shared congestion window's utilization.These contributions collectively represent a comprehensive effort to address the challenges of vehicular network communications through the QUIC protocol, and can be visualized in Table I.By replacing the QUIC congestion control algorithm with our (now further enhanced) S-HTCP, and by improving the flow control mechanism of QUIC, our work lays the groundwork for more reliable, efficient, and fair communication in highly dynamic network environments.
III. BACKGROUND
The QUIC protocol offers several features that enhance its functionality and make it suitable for various communication scenarios.One key feature is its ability to multiplex different streams over a single UDP connection, allowing for concurrent and independent data transmission.QUIC also significantly reduces connection establishment latency, with a best-case scenario of one RTT and the potential for zero-RTT connection establishment (when the client has interacted with the server before), which significantly outperforms traditional TCP-based connections.QUIC ensures the security of data delivery through authenticated and encrypted header and payload.QUIC It offers diverse flow control mechanisms at both the connection and stream levels, allowing the sender to adjust the amount of data transmitted based on the receiver's advertised capacity [7].This efficient flow control prevents overwhelming the receiver and ensures smooth data transmission.QUIC provides flexible congestion control mechanisms, including the default CUBIC algorithm as well as that of Bottleneck Bandwidth and RTT (BBR) [6].It also employs a packet pacing mechanism to effectively manage data bursts and handle network traffic.Additionally, QUIC supports connection migration, which is particularly useful in scenarios involving IP address changes.By using a 64-bit connection ID and maintaining the same session key, QUIC ensures seamless connection maintenance and allows migrating clients to maintain authentication and cryptographic verification.These features collectively contribute to the efficiency, security, and adaptability of the QUIC protocol, making it a promising choice for various communication scenarios, including vehicular communications [18].In this article, we will delve into the fundamental mechanisms and operational principles of the QUIC protocol, with a specific focus on the congestion and flow control control mechanisms.The ultimate goal of this work is to propose a set of channel-aware mechanisms for QUIC, in order to facilitate the secure, timely and reliable exchange of CAMs.
A. QUIC: Congestion Control
Congestion control is a critical component of any transport protocol, including QUIC.It aims to ensure that the amount of data sent by a sender does not exceed the capacity of the network and avoids congestion collapse.Most QUIC implementations employ a variant of the CUBIC congestion control algorithm to regulate the sending rate and adjust the congestion window size.CUBIC is a popular congestion control algorithm that aims to provide a scalable, stable, and fair mechanism for managing congestion in the network.It utilizes a cubic function to estimate the available network capacity and adaptively adjust the sending rate accordingly.However, it's important to note that CUBIC is primarily used for connection-level congestion control in QUIC, rather than stream-level congestion control.There also exist implementations leveraging the BBR algorithm, which probes the network to accurately estimate the available bandwidth and delay, and then adjust the sending rate accordingly.BBR's main advantage is its capability to fully use bandwidth, despite high packet losses [19].
B. QUIC: Flow Control
In QUIC, each stream within a connection has its own flow control mechanisms.Flow control ensures that a receiver does not get overwhelmed by data from a sender, while congestion control regulates the sending rate to avoid network congestion.While QUIC provides stream-level flow control, which dynamically adjusts the receive window for each stream based on available buffer space, it currently lacks a mechanism for reallocating the congestion window in case of delayed ACKs at the stream level.In the absence of a re-allocation mechanism, if a stream does not receive timely ACKs for the data it has sent, the neighbouring streams and the overlaying connection may not be able to fully utilize their allocated congestion window.This can lead to under-utilization of network-wide resources and sub-optimal performance, especially for streams experiencing delays in receiving ACKs.This will be discussed in greater detail in IV-B.
C. Cooperative Awareness Messages (CAMs)
In our work, we propose the utilization of QUIC as a transport-layer protocol, the payload of which are CAMs.Those messages are essential for facilitating effective communication among vehicles and infrastructure in intelligent transportation systems -they provide crucial information about a vehicle's state, enabling cooperative functionalities (e.g., collision avoidance and traffic management).CAMs are broadcasted periodically, at a given frequency defined by application requirements and network limitations [20].Currently, CAMs are traditionally transmitted over UDP due to its real-time capabilities and overall suitability for real-time applications.However, UDP lacks any form of reliability mechanisms, necessitating additional error detection and recovery at the application layer; using our modified version of QUIC as a transport-layer protocol enables built-in reliability, congestion control, and security, improving the efficiency and integrity of CAM transmission.This development is in aligned with the goals of O-RAN, promoting seamless and secure communication between vehicles and infrastructure.
IV. METHODOLOGY -ENHANCING THE QUIC PROTOCOL
Our proposed enhancement of the QUIC protocol is twofold.Firstly, we propose the utilization of the S-HTCP congestion control algorithm which we proposed in [2], which has been further enhanced to consider the received singular strength indicator (RSSI) metric.Secondly, we propose a new methodology to allocate the available congestion window of a given QUIC connection's individual streams.
A. QUIC -Congestion Control Enhancement
By default, QUIC uses either CUBIC or BBR as the default congestion control algorithm.CUBIC, which is a heuristic algorithm driven by sender-side events, utilizes real-time scaling metrics instead of RTT-based metrics.BBR considers channel characteristics and parameters to adjust pacing rate and congestion window gain, starting from a given value.Equations 1 [21] concerns the Pacing Rate parameter (rate at which packets are sent), while Equation 2[21] concerns model the way in which BBR implements channel awareness in practice, considering that at a given time t in a considered period T , we can define BtlBw as seen in Equation 2. Pacing = min(Gain • BtlBw, max(Pacing)) (1) where: Gain = Pacing scaling factor, determines pacing rate BtlBw = Estimated bottleneck bandwidth W b = Bandwidth filter window T = Time period considered by BRR Similarly, the congestion window calculation process in CUBIC can be seen in Equation 3 [7], and is a direct function of the K time offset parameter.Said offset parameter represents the time when the cubic curve started.It is calculated as the third root of the product of the maximum window size (W m ax) and the complement of the beta parameter (1 − beta), divided by the scaling constant (C).The value of K essentially sets a baseline for the congestion window's growth and reduction.It influences how quickly the congestion window increases during the additive increase phase and how aggressively it decreases during the multiplicative decrease phase in response to congestion signals.Equation 4 correspondingly describes this metric.
where: C = Scaling constant determining aggressiveness T = Time elapsed since the last cwnd reduction K = Offset parameter depicting the cubic curve start w max = Window size before the last reduction β = Decrease factor of the AIMD algorithm 1) S-HTCP: Our Congestion Control algorithm: In our previous work in [2] we proposed a new set metrics for the AIMD algorithm of the H-TCP protocol and managed to achieve an improvement in total throughput and end-toend delay.D. J. Leith et al. in [22] define the α and β factors which we used to formulate our own algorithm variant.Equations 5 and 6 show the baseline of our two previously proposed AIMD metrics.Note that we are assuming that TCP has crossed the threshold for switching from standard TCP operation to the new increase function.Normally, the second component of the product shown in Equation 5 would be expressed as 1 + 10(∆ 2 , where ∆ L is the aforementioned threshold after which the protocol switches to the operation of interest.In our case, we disregard the previous operational phase and thus replace ∆ L with a null value.Simplifying the expression yields ∆ 2 /4 + 10∆ + 1. β S−HT CP = e −∆/λ1 RT T min RT T max (6) where: In this case, assuming a successful acknowledgement, the congestion window is defined as cwnd ← cwnd + α/cwnd, while in the case of congestion event cwnd ← β • cwnd.
2) Enhancing S-HTCP using RSSI Volatility Parameters: In the context of this work we have additionally enhanced the S-HTCP AIMD algorithm to consider the volatility of the RSSI value, as well as the direction thereof.In order to capture the behaviour of the links in the same medium through the RSSI, we follow a coherent methodology described below.The methodology has been implemented in NS-3 and can be adopted by public implementations of the QUIC protocol with relative ease and minimal overhead.In addition to considering the time elapsed since the last congestion event and the min/max RTT, we do the following: First we calculate the RSSI difference (∆RSSI) between consecutive measurements: where RSSI[n] represents the RSSI value at time step n and RSSI[n − 1] represents the RSSI value at the previous time step.Secondly, we calculate the time difference (∆t) between consecutive measurements: where t[n] represents the time at time step n and t[n − 1] represents the time at the previous time step.Thirdly, we calculate the rate of change of RSSI (RSSI slope) using the derivative formula: RSSI slope = ∆RSSI/∆t, outputting the rate at which the RSSI is changing per unit time.Fourthly, we apply smoothing to the RSSI slope to reduce noise or fluctuations.Our approach is to calculate the cumulative average (CA): Smoothed RSSI slope = CA(RSSI) = where k is the number of previous RSSI slope values to be considered in the moving average calculation.The eventually computed metric is utilized by means of sigmoid scaling for the α AIMD parameter, considering inverse scaling for the β expression.
What we have thus achieved is the fact that the protocol can now re-adjust the additive increase rate to be greater for connections with a larger congestion window.By scaling the α parameter accordingly, this approach offers resilience against abrupt RTT changes, similarly with the observed convergence time -which in turn further reduces RTT unfairness between competing flows.
Using this information, we can re-write the expressions in Equations 5 and 6 as follows in Equations 7 and 8, representing the enhanced α and β parameters previously analyzed.
RT T min RT T max (10) where: The inclusion of RSSI volatility as a parameter in our extended AIMD algorithm provides valuable insights into the stability and quality of wireless connections.By monitoring RSSI fluctuations, our approach can dynamically adjust the congestion control strategy based on the reliability of the wireless link.This adaptability allows a connection's congestion window to respond more effectively to potential network congestion or interference, thereby ensuring uninterrupted data transmission and mitigating performance degradation.Additionally, the previously proposed consideration of the time elapsed since the last congestion event, offers a dynamic perspective on network stability.By incorporating this parameter, the algorithm intelligently adapts the above-described AIMD parameters based on the historical behavior of the network.This enables QUIC's congestion control mechanism to avoid unnecessary CWND reductions during transient congestion periods, resulting in improved network efficiency and faster recovery after congestion events.
In the context of improving the reliability and lowering the latency of autonomous driving and V2X communications, our proposed approach of considering RSSI volatility instead of Signal-to-Noise Ratio (SNR) for congestion window optimization in O-RAN and its integration with C-V2X technology offers significant advantages: reliability and low latency are crucial for C-V2X systems, where real-time and reliable communication between vehicles, networks, and infrastructure is essential.By incorporating RSSI volatility, our approach captures the dynamic changes in the RSSI, providing a comprehensive understanding of the wireless link's stability.Unlike SNR, which offers a static measure of signal quality using Rx power estimations [23], RSSI volatility reacts swiftly to sudden variations in the link, enabling proactive congestion control measures.By accurately monitoring and considering RSSI volatility, our algorithm facilitates precise congestion window adjustments that reflect real-time network conditions.This enables quicker response times and reduces CAM delays.The ability to make informed congestion control decisions based on nuanced RSSI variations enhances the reliability and responsiveness of autonomous driving systems.To further elaborate on the reasoning behind choosing RSSI over SNR volatility as a metric, focusing on RSSI aligns our work with prevalent practices in real-world cellular and vehicular communication systems, where this metric is commonly used for various operational decisions e.g., adaptive rate control or (cellular) handover mechanisms.This choice is poised to enhance the practical applicability of our findings to (current and future) C-V2X technologies.Additionally, given that RSSI is noise-model agnostic, it provides a valuable measure of link quality volatility without the need for detailed noise level estimation.This approach simplifies the simulation setup while still allowing for accurate modeling of signal propagation and reception under varying conditions.This ensures the broad applicability of our research findings across diverse vehicular and wireless scenarios, not limited by specific characteristics of the noise environment or detailed signal quality assessments.
It is important to note that we differentiate between the effects of fast fading and large-scale fluctuations on RSSI measurements.Fast fading, characterized by rapid, short-term changes in signal strength due to multipath scattering, presents a challenge to maintaining stable communication channels.To mitigate its impact and prevent the algorithm from reacting to these transient variations, we employ sophisticated smoothing techniques on RSSI values as discussed previously which "cancel" out the effects of fast fading.This approach ensures that our congestion control adjustments are based on more stable trends in signal strength, primarily influenced by factors such as shadowing and the mobility of nodes instead of .This distinction enables it to make informed decisions about congestion window adjustments, thereby enhancing overall reliability.Summarizing, the deliberate application of smoothing functions to RSSI measurements allows us to extract meaningful trends from the link quality volatility whilst minimizing the effects of sudden RSSI drops.
B. QUIC -Dynamic Flow Control and Resource Allocation Optimization
QUIC inherently creates several streams per connection.At the connection level, QUIC uses CUBIC or BBR to avoid congestion.At the stream level, QUIC uses a limit-based flow control mechanism.A receiver advertises the limit of total bytes it is prepared to receive on a given stream or for the entire connection.Stream flow control attempts to prevent a single stream from monopolizing the entire receive buffer.It also prevents senders from exceeding the buffer capacity of a receiver -this is done by limiting the total bytes of stream data sent in STREAM frames.For each stream, QUIC maintains a sending rate to ensure fair sharing of available bandwidth among streams within a connection.It regulates the sending rate based on feedback received from the receiver, such as ACKs and information about the available buffer space at the receiver.This approach allows QUIC to adapt to varying network conditions, handle congestion effectively, and deliver improved performance for real-time applications, which is of particular importance in low-latency C-V2X environments.
The interconnection between streams' congestion control and flow control mechanisms in QUIC plays a crucial role in ensuring efficient and reliable data transmission.The congestion control mechanism regulates the rate at which data is sent over the connection, while the flow control mechanism manages the amount of data that can be sent on individual streams.These two mechanisms work together to prevent congestion and ensure fair resource allocation within the QUIC protocol.At the stream level, flow control prevents a single stream from monopolizing the receive buffer by limiting the amount of data that can be sent on each stream.The receiver advertises the initial flow control limits for all streams during the handshake, and subsequently sends MAX_STREAM_DATA frames to increase the limits.This mechanism allows the receiver to control the rate at which data is received on each stream, ensuring that a stream does not consume more resources than allocated.
At the connection level QUIC maintains a congestion window, which represents the allowed number of packets in flight at any given time, which is shared among all streams.This ensures that the overall transmission rate does not overwhelm the network and prevents congestion.The interaction between congestion control and flow control is evident in the way ACKs affect the allocated proportion of the connection CWND for each stream.When a stream receives acknowledgments, the flow control window is updated to increase the allowed transmission rate for that stream.This allows well-performing streams to utilize more of the available bandwidth.However, if a stream does not receive acknowledgments, its allocated proportion of the connection CWND (supplied to the stream in the form of the flow control window) remains unchanged, ensuring fairness among streams and preventing a poorly performing stream from consuming a larger share of the available resources.
The existing state of flow control in QUIC presents a profound challenge wherein an individual stream, due to a diminished acknowledgment rate, may result in sub-utilization of its allocated flow control and congestion window.This, in turn, creates a bottleneck situation as the remaining streams within the same congestion group currently lack the capability to dynamically update their congestion windows and adjust sending rates in response to the reduced ACK rate of the affected stream.To overcome this limitation, we propose a refined approach to the stream-level flow control resource management system within QUIC.
1) Consideration of Connection-level Global Optimum: Our proposition involves considering the global optimum of the connection, instead of treating each stream in isolation.By empowering the other streams to adapt their windows and sending rates based on feedback received from the affected stream, we can effectively mitigate the bottleneck caused by the under-utilization of the allocated window.Implementing this modification necessitates enhancements to the existing flow control mechanisms of QUIC.It entails establishing robust communication and coordination mechanisms among streams to facilitate the exchange of ACK-related information and enable collective decision-making regarding sending window adjustments.By incorporating insights from all streams within the connection, we can achieve a harmonized allocation of network resources and optimize the overall sending rate of the connection.This approach not only mitigates the bottleneck effect induced by a single stream but also maximizes the utilization of the shared congestion window, thereby elevating performance and fairness across all streams within the connection.Currently, a QUIC sender is set to ignore any MAX_STREAM_DATA or MAX_DATA frames that do not increase flow control limits.[1].Effectively, a stream has consumed its allocated flow control limit (which is a function of the connection's shared resources, the number and type of concurrent streams, as well as the prioritization thereof), it will be blocked from increasing its sending rate.If a stream (usually of a higher priority) fails to timely receive ACKs, its allocated window will still occupy the shared resources and will reduce connection-wide throughput.Varying ACK rates in QUIC streams will generally result in this connection-wide resource under-utilization.
Algorithm 1 Data Transmission Function
Require: Stream data block status Ensure: Singular stream data transmission 1: while data to be sent do if StreamBlocked (stream) then if DataBlocked() then 7: DAT A BLOCKED(f rame) for each streamID in Connection do 3: lastAckT ime ← LastAckT ime(streamID)
9:
ST REAM DAT A F ACT OR) 10:
return adjustedW indow
Considering the above, we are proposing the addition of the DynamicFlowControl function in the QUIC flow control mechanism.Algorithms 1, 2 and 3 offer a detailed view of how the proposed enhancement is implemented in practice, while Figure 2 gives a high-level overview of our proposed function.The proposed function is to be called within the FlowControlAdjustment routine within the same mechanism.The algorithm consists of three main functions: DataTransmission, FlowControlAdjustment, and DynamicFlowControl.These functions work together to enable data transmission and dynamically adjust flow control parameters.The DataTransmission function handles the process of sending data packets over the network.It operates in a loop, continuously checking if there is data to be sent.If the flow control window for a specific stream (StreamBlocked(stream)) is full, indicating that the receiver cannot accept more data, the function notifies the receiver by sending a STREAM_DATA_BLOCKED frame.This action halts the data transmission temporarily, allowing for flow control adjustment.FlowControlAdjustment is then called to adjust the flow control parameters based on received frames.Upon receiving a MAX_DATA frame, which indicates the maximum flow control window size at the connection level, the algorithm updates the flow control window size (FC_WND) accordingly.If the flow control window is still blocked (DataBlocked()), meaning that the receiver cannot accommodate more data, a DATA_BLOCKED frame is sent to notify the receiver about the limitation.When a MAX_STREAM_DATA frame is received, representing the maximum flow control window size for a specific stream, the algorithm updates the flow control window size (FC_WND) for that stream using the DynamicFlowControl function which calculates the adjusted window size based on the elapsed time since the last acknowledgment received for the streams of the connection at hand.If the flow control window is blocked (StreamBlocked(frame.streamID)), indicating that the receiver cannot receive more data for that stream, a STREAM_DATA_BLOCKED frame is sent to inform the receiver.Similarly, upon receiving a MAX_STREAMS frame, which specifies the maximum number of streams allowed for a given stream type, the algorithm updates the maximum allowed streams (AllowedStreams max ) accordingly.If the stream limit is reached (StreamBlocked(stream)), indicating that no more streams can be created, a STREAMS_BLOCKED frame is sent to notify the receiver.The DynamicFlowControl function is responsible for calculating the adjusted window size for stream-level flow control.It iterates over each stream in the connection and determines the elapsed time since the last acknowledgment (LastAckTime(streamID)) for each stream.The elapsed time is then used to calculate the adjusted window size by multiplying it with a predefined STREAM_DATA_FACTOR and adding it to the base window size (MAX_STREAM_DATA_BASE).This dynamic adjustment takes into account the varying network conditions and the receiver's ability to handle data.
V. EVALUATION
This section is dedicated to the analysis of the evaluation of our proposed enhancements to the discussed QUIC mechanisms.It entails an analysis of our simulation environment, considered mobility models, the structured of the exchanged benchmark messages, as well as the actual results.A total of two emulation scenarios have been designed, each based on a combination of different mobility models for aerial and terrestrial nodes.Our simulation environment is built around several NS-3 modules as well as additional software such as Simulation of Urban Mobility (SUMO) [24], specialized in modelling vehicular mobility in emulated real-world environments.
A. Simulation Environment
Regarding our utilized NS3 modules, firstly, F. Eckermann et al. in [4] introduce an NS-3 based C-V2X simulator which constituted the foundation of our emulation framework.
Secondly, we used the QUIC implementation in NS-3 offered by A. De Basio et al. in [16].The implementation (available in GitHub [25]) is aligned with version 13 of the IETF QUIC drafts and is based on the NS-3 TCP implementation and includes improved acknowledgement mechanism, multiplexing of different streams in a single connection, 0-RTT handshake, possibility for custom stream schedulers, as well as BBR which is of utmost importance for our application.
Thirdly, our simulation environment consist of the O-RAN E2 interface NS-3 module [3].If our case, it is used to facilitate the exchange of CAMs between the RAN infrastructure and the connected vehicle systems, through aerial relays.Table II offers an overview of the most important simulation parameters.
Fourthly, as already mentioned, for one of the considered scenarios, we have utilized SUMO to model an accurate and realistic vehicular traffic scenario, based on the TAPAS-Cologne [26] scenario.
B. Vehicular CAMs
Each exchanged CAM message is constructed as a string and converted to a packet for transmission.The message format includes the fields described in III.Received CAM messages are logged, along with other relevant information to separate CSV files.The logged information includes the transmitted and received CAM messages themselves, simulation time and statistics (simulation time, total received rackets and total transmitted packets. Algorithm 4 provides a high-level overview of how CAMs are generated.Specifically, the algorithm takes a list of vehicles as input and generates a list of CAM messages as return camMessagesList C. Mobility Models a) Scenario 1: Linear Vehicular Mobility: For this scenario, we consider that terrestrial vehicles' mobility vectors are lineal and can be described by the constant acceleration mobility model.
Aerial ad hoc nodes are tasked with implementing cellular communication relaying, leveraging the anchored selfsimilar 3D Gauss-Markov mobility model which we proposed in [17].This mobility model is geared towards modelling communication-relaying scenarios leveraging aerial nodes and updates the process of calculating the new velocity of nodes by introducing the relative velocity between two directly associated nodes as a weighted and exponentially decaying positional index.The observed outcome is a more fluid and stabilised relative acceleration, which leads to a more positionoriented deployment of the swarm, a key enabler in communications relaying.Equation 11 [17] describes the process of setting the new speed for a node using a randomness index which has been enhanced to accurately model communicationsrelaying applications.Similarly, Equation 12 [17] models the process of setting the new direction, while Equation 13 [17] mathematically models the process of assigning a new pitch value for a node.The foundation of all three expressions is the Gauss-Markov mobility model which has been modified as we document in detail [17].V xyz , which is the mobility vector in 3D space of a given networked entity, where for the mobility vector component projected in the x axis it is true that ⃗ V x = (s n cos(d n ) cos(p n )) î, while regarding the component projected in the y axis we have ⃗ V y = (s n sin(d n ) cos(p n )) ĵ, and for the component projected in the z axis we have ⃗ V z = (s n sin(p n )) k.The applied technique results in spatial "anchoring" in regards to nodes' previous speed and direction respectively.Effectively, this smooths the rate at which direction changes whilst also maintaining the Gaussian randomisation attribute of the overarching positional mechanism.
b) Scenario 2: TAPASCologne-based Mobility: This scenario constitutes an additional, highly realistic hybrid NS3-SUMO-based scenario, considering actual vehicular mobility as derived from TAPASCologne.It provides a comprehensive simulation environment designed to model a realistic vehicular traffic within Cologne, Germany, across a day.The original data, initially aligned with a proprietary road network, has been meticulously mapped to a network derived from OpenStreetMap.The scenario package includes road networks from OSM, Points of Interest (POIs), polygons, and mapped trips covering early morning to late evening hours.In our application, we focused on a subset of the modelled vehicles, This scenario can be considered hybrid in terms of mobility, as it utilizes two sources for the definition of node mobility vectors.Specifically, all aerial nodes are assigned the same anchored self-similar Gauss-Markov-based mobility model, as described in the previous scenario.However the initial locations of the aerial nodes are defined as a function of the outline of the locations given by the vehicular nodes.Regarding the vehicular nodes, their mobility parameters are not given by a mobility model as before, but are rather extracted from TAPASCologne.
To facilitate the integration of SUMO-derived mobility models into NS-3, we employed the NS3 traceExporter tool to export SUMO mobility traces into NS-2 trace format, suitable for NS-3 consumption.This conversion is pivotal for harnessing realistic vehicular mobility patterns within network simulations.After running the simulation with the desired parameters, we generate the NS-2 trace file (simulation.ns2.tr)encapsulating vehicle movements.Subsequently, within NS-3, the Ns2MobilityHelper class reads and applies these mobility patterns to network nodes directly from the .trfile.This approach replaces mobility models for vehicular nodes.
D. Results
The conducted experiments considered a use case where mobile vehicular nodes unicasted streams of CAM data to a networked entity, over aerial nodes functioning as mobile relays.The results constitute averaged values of numerous experiments and clearly demonstrate that the proposed enhancements to the QUIC protocol can yield tangible performance increases in a C-V2X scenario.Our end-goal is to practically show that; 1) The proposed enhancement of the S-HTCP congestion control algorithm, considering RSSI volatility on top of the ratio between minimum and maximum RTT and time elapsed since the last congestion event, manages to increase total throughput whilst maintaining substantially lower latency when coupled with native QUIC implementations.This statement should be valid while the relative velocity (as measured between mobile networked entities and their corresponding relays) is increased.
2) The proposed enhancement of the native QUIC flow control mechanism, considering an active connection's streams' time elapsed since the last received ACK will yield increased total throughput per connection as the offered load increases.Again, as the relative velocity (as measured between mobile networked entities and their corresponding relays) is increased, total throughput should be less affected.
In terms of achieved communication quality, we consider the relationship between offered load and RTT, as well as that between offered load and connection throughput.Figure 4 visualizes the performance of the benchmarked QUIC variants with the three examined congestion control algorithms (CUBIC, BBR, and Modified S-HTCP, the last of which was also enhanced with the proposed dynamic flow control mechanism) in terms of RTT as the offered load is increased.It becomes evident that the variant of Modified S-HTCP Congestion Control w/ Dynamic Flow Control measurably outperforms the other two QUIC implementations based on CUBIC and BBR respectively.Our proposed mechanisms enable QUIC to achieve lower round trip times as the offered load increases, and does so in a less volatile and more deterministic manner.Regarding the CUBIC-based QUIC variant, it can be observed that despite initially having better performance compared to BBR-based QUIC, its behaviour is more erratic.Regarding the BBR-based QUIC implementation, we observe a behaviour similar to that of our proposed implementation, though with a decreased performance and reduced linearity.This performance increase can mainly be attributed to the proposed S-HTCP-based congestion control algorithm.
The measured RTT values are averages achieved after multiple simulations per scenario, per protocol/algorithm combination.As to the differences in trends observed between the contender protocols, those can be mainly attributed to the impact of congestion control algorithms (varying responses to network condition changes), network dynamics and the impact of relative velocity.Considering the rate at which RTT increases as we adjust the offered load, we can deduct that our proposed protocol and algorithm combination achieves a statistically significant lower end-to-end delay.In practice, RTT is measured by utilizing simple acknowledgements and measuring the total time elapsed for such an event.Lastly, we need to remark that our implementation of the congestion control algorithm utilizes by itself the measured end-to-end delay (both the minimum and maximum values, as well as the ratio thereof) as a metric to dynamically adjust sending rates.Therefore, it is expected that when offered greater loads, the variant of the protocol capable of converging faster to the optimal operating point will yield superior results in terms of RTT.
Continuing, Figure 5 visualizes the performance of the same QUIC variants in terms of throughput as the offered load is increased.Again, our proposed enhancements to the QUIC congestion and flow control mechanisms outperform the native CUBIC-and BBR-based QUIC benchmarks.Specifically, our proposed mechanism enhancements enable QUIC to achieve higher throughput as the offered load increases, which can be attributed to the way in which data streams now implement flow control and calculate their corresponding windows.While minimal, the consistent increase in throughput of our proposed mechanisms indicate that C-V2X applications would greatly benefit from similar dynamic flow-control implementations.
As highlighted in [27], dynamic propagation delay introduced by higher degrees of mobility increases the possibility of packet loss.By default, transport-layer protocols consider all packet loss events as indications that the network is experiencing congestion.Thus, the sending rate accordingly is appropriately reduced.In our evaluation, we thus considered RTT and throughput performance as a function of propagation delay volatility, expressed through relative node velocity.incorporating the proposed Dynamic Flow Control mechanism follows a clearly linear behaviour with a high degree of determinism while the other variants exhibit greater variance coupled with under-performance.
Figure 7 illustrates the performance of the three QUIC variants in terms of throughput as the relative node velocity is increased.Assuming a completely static network, the BBRbased variant demonstrates increased performance.However, as relative node velocity is increased BBR's performance is shown to greatly deteriorate, and is surpassed by both the CUBIC-based implementation and our own proposed mechanisms.More specifically, the S-HTCP-based variant coupled with Dynamic Flow Control appears to be comparatively less affected by changes in relative node velocity.This also translates to increased resilience of the network against volatile propagation delay which is by definition a common occurrence in vehicular networks in C-V2X environments.
Figure 8 illustrates the achieved end-to-end delay as a function of the total network size.The second mobility scenario of those described in Subsection V-C was used to generate these results, involving up to a total of 28 nodes (including the deployed aerial relays).In order to conduct this experiment, each set of (a maximum of) 6 vehicles was connected to a network of UAVs, and different point-to-point paths were established to benchmark the network.Considering the nature of the deployment (half duplex CAM unicasting) and the negligible propagation delay, the end-to-end delay is generally in acceptable levels (less than 1ms in almost all cases) even considering repeated relaying and the channel propagation and shadowing models.Our proposed variant seems to outperform the native QUIC implementations in terms of scaling better as the network becomes more congested.
VI. CONCLUSIONS
We have considered highly advanced C-V2X scenarios incorporating aerial relays, engaging in realistic mobility using both custom relevant models, as well as actual street map data derived from SUMO, as described in Subsection V-C.In our scenario, vehicular nodes exchange CAMs which are implemented as custom frames in NS-3, over established QUIC connections.We propose a set of enhancements for the implementation of QUIC, utilizing a novel congestion control mechanism and a global-optimum aware dynamic flow control algorithm.Our evaluation proves that enabling channelawareness in two core mechanisms of the QUIC protocol yields increased efficiency in C-V2X deployments over aerial base stations.
To further validate our findings, we considered an anchored self-similar Gauss-Markov-based mobility model for aerial nodes, capable of accurately modelling the behaviour of drones engaging in communication relaying.Coupled with dual mobility scenarios for vehicular nodes, our work presents a strong case in terms of realism, both from a physical and a network perspective.Similarly, we consider the impact of relative speed on throughput and RTT by introducing constant acceleration to the corresponding vehicular entities' models, as mobility is a critical consideration for mobile ad hoc networks [27].
Our implementation demonstrates a high degree of resilience against speed volatility, hinting that such improvements can bring about the realization of truly advanced autonomous vehicular services networked.The proposed enhancements are relatively easy to implement in NS-3 and are compatible with existing QUIC implementations, making our proposed work a viable solution for improving the performance of QUIC in transient networks.
Overall, our research paves the way for the next generation of vehicular communication technologies bringing about reliable yet UDP-like fast communications for the control and user planes alike, taking into account future developments in the domain of ad hoc networking and vehicular mobility.By meticulously addressing the intricacies of V2X networks, including the integration of aerial relays and the adoption of sophisticated mobility models, we underscore the potential for QUIC protocol adaptations to significantly elevate network performance.The introduction of our novel congestion control and dynamic flow control algorithms not only showcases a marked improvement in network efficiency but also emphasizes the protocol's adaptability to the fluctuating dynamics and size of vehicular networks and their channel parameters.This adaptability is crucial for supporting the diverse and evolving needs of modern vehicular applications, from autonomous driving to real-time traffic management systems and crisis mitigation endeavours.Furthermore, our approach demonstrates a harmonious balance between theoretical innovation and practical applicability, offering a scalable and forward-thinking solution for future vehicular communications and heralding a new era of mobility and connectivity. ,time...
Figure 1 .
Figure 1.Envisioned system model Time since the last congestion event RT T min = Minimum round trip time RT T max = Maximum round trip time λ1 = Exponential decay constant 1 λ2 = Exponential decay constant 2 Time since the last congestion event RT T min = Minimum round trip time RT T max = Maximum round trip time λ = Simplified exponential decay constant CA(RSSI) = Cumulative average of a connection's RSSI
Figure 4 .
Figure 4. Achieved RTT as a function of offered load
Figure 5 .
Figure 5. Achieved throughput as a function of offered load
Figure 6 .
Figure 6.Achieved RTT as a function of relative node velocity
Figure 7 .
Figure 7. Achieved throughput as a function of relative node velocity to-End Delay (ms) Average End-to-End Delay vs. Number of Nodes QUIC -CUBIC Congestion Control QUIC -BBR Congestion Control QUIC -Modified S-HTCP w/ Dynamic Flow Control
Figure 8 .
Figure 8. End-to-end delay as a function of network size It processes each vehicle in the input list individually, following a series of steps.Firstly, it creates a new CAM message object.Then, it assigns the Vehicle ID attribute to the ID of the current vehicle.Subsequently, the algorithm sets the Simulation Time attribute to the specified simulation time.Next, it sets the Position (x-coordinate) attribute to the x-coordinate of the current vehicle's position, and the Position (y-coordinate) attribute to the y-coordinate of the current vehicle's position.After completing these attribute assignments, the algorithm adds the generated CAM message to the list of CAM messages.Finally, the algorithm returns the complete list of generated CAM messages. 10: | 11,793.6 | 2024-07-01T00:00:00.000 | [
"Engineering",
"Computer Science",
"Environmental Science"
] |
A Study on the Influence of Tire Speed and Pressure on Measurement Parameters Obtained from High-Speed Tire Uniformity Testing
: In order to improve the test conditions of the tire uniformity test and the e ff ect of the speed and tire pressure on the uniformity parameters, the uniformity test of the tire under di ff erent speeds and tire pressure was carried out by a high-speed uniformity test machine, and the experimental data were analyzed and fitted by the regression analysis method. This paper introduces the definition of uniformity and the uniformity parameters of automotive tires; the working principle of a high-speed uniformity testing machine is briefly described, a mathematical model of the uniformity testing machine is established, and the signal acquisition process of the tire uniformity parameters and the calculation method of the uniformity parameters are described. The test result indicates: As the speed increases, the radial force fluctuation, lateral force fluctuation, tangential force fluctuation, and turning torque fluctuation of the tire increase, and the positive torque fluctuation first increases and then decreases; with the increase of tire pressure, the radial force fluctuation and the tangential force fluctuation of the tire increase, and the lateral force fluctuation, the turning torque fluctuation, and the returning moment fluctuation are all reduced. Compared to the low speed uniformity test, the high speed uniformity test can better reflect the uniformity of the tire, reducing the speed of the vehicle can reduce the radial runout and lateral sway of the tire; increasing the tire pressure can reduce the left and right swing of the vehicle.
Introduction
With the rapid development of the automotive industry, the performance of cars in the past no longer meets the needs of drivers in the present; vehicles are constantly improving in their handling and comfort. As the only part of the vehicle in contact with the road surface, the performance of the tires directly affects the performance of the whole vehicle. Therefore, the automotive industry pays attention to and cooperates with the tire industry as part of its efforts to improve the performance of vehicles, so that vehicle stability and comfort are further enhanced.
A tire will deform irregularly when driven at high speed, resulting in fluctuations in radial, lateral, and tangential forces, reflecting variation in the circumferential forces of the tire, known as tire nonuniformity. Due to the unevenness of the tire, the forces acting on the rotating shaft during the rolling process cause the vehicle to vibrate, which reduces the maneuverability, ride, and comfort of the car, and can seriously damage some of the car's components. Tire uniformity testing is a major part of tire quality testing. For newly produced tires, uniformity testing is required to make sure that the actual measurement uniformity parameter error is within the allowable tolerance range, ensuring the tire quality assurance during use [1]. The high-speed uniformity testing machine is a device that detects the fluctuation of the circumferential forces of the tire by driving a drum to rotate the tire and applying a certain load to the tire and can accurately measure the tire uniformity parameters, thereby being able to characterize the performance of the vehicle. The high-speed uniformity testing machine is a semi-automatic testing device for measuring tires under different speeds, loads, and pressures. The uniformity parameters of the test include radial force fluctuation, lateral force fluctuation, tangential force fluctuation, radial force harmonics, lateral forces harmonics, tangential force harmonics, positive torque fluctuations, rollover torque fluctuations, taper effects, and angle effects [2]. According to the measured test data, not only can the quality of the tire be judged, but also the performance of the tire can be improved by using the uniformity parameters to improve the performance of the vehicle.
In the past, many scholars mostly used the low-speed uniformity testing machine to study the uniformity parameters of tires, and used it as the standard for tire quality evaluation [3]. At present, the speed of vehicles on the highway is about 120 km/h, and the test results of the low-speed uniformity tester cannot accurately reflect the forces of the tire when it is rolling at high speed; some parameters that are neglected in the low-speed uniformity test may be important reference factors in high-speed uniformity tests, such as tangential force fluctuations, positive-torque fluctuations, and roll-over torque fluctuations. The accuracy of the sensors of the low-speed uniformity tester does not meet the test requirements of these parameters, and does not fully reflect the uniformity characteristics of the tires; these parameters can be used to characterize the uniformity of the tires at high speed. In a tire uniformity test, there are many factors that affect the unevenness of the tire, such as speed, load, and pressure, which should be examined in order to study the change of forces and moments when the tire is rolling at high speed. In this paper, a high-speed uniformity tester is used to measure the change of forces and moments of a tire under different speeds and different pressures, and the influence of speed and pressure on tire uniformity parameters is analyzed. Additionally, the relationship between tire uniformity parameters and vehicle performance is briefly described, providing a reference for improving the production level of tires and vehicle design.
Tire Uniformity Definition
A tire is an annular elastic body made of various materials such as rubber, steel cord, and polyester. Its quality, size, and rigidity are uneven during the structural design and production process. Tire uniformity testing means that the tire is inspected for uneven size, mass, and force under certain conditions of pressure, load, and speed [4]. Tire uniformity measurement involves the rotation of the tire by applying a load on a rotating drum that is in contact with the tire. Force sensors mounted beside the tire shaft measures the fluctuation of forces and moments in the circumferential direction. The unevenness of the tire mainly includes uneven mass, uneven size, and uneven rigidity [5]. Figure 1 shows a schematic of the unevenness of a tire.
Vehicles 2020, 2, FOR PEER REVIEW 2 of 14 part of tire quality testing. For newly produced tires, uniformity testing is required to make sure that the actual measurement uniformity parameter error is within the allowable tolerance range, ensuring the tire quality assurance during use [1]. The high-speed uniformity testing machine is a device that detects the fluctuation of the circumferential forces of the tire by driving a drum to rotate the tire and applying a certain load to the tire and can accurately measure the tire uniformity parameters, thereby being able to characterize the performance of the vehicle. The high-speed uniformity testing machine is a semi-automatic testing device for measuring tires under different speeds, loads, and pressures. The uniformity parameters of the test include radial force fluctuation, lateral force fluctuation, tangential force fluctuation, radial force harmonics, lateral forces harmonics, tangential force harmonics, positive torque fluctuations, rollover torque fluctuations, taper effects, and angle effects [2]. According to the measured test data, not only can the quality of the tire be judged, but also the performance of the tire can be improved by using the uniformity parameters to improve the performance of the vehicle.
In the past, many scholars mostly used the low-speed uniformity testing machine to study the uniformity parameters of tires, and used it as the standard for tire quality evaluation [3]. At present, the speed of vehicles on the highway is about 120 km/h, and the test results of the low-speed uniformity tester cannot accurately reflect the forces of the tire when it is rolling at high speed; some parameters that are neglected in the low-speed uniformity test may be important reference factors in high-speed uniformity tests, such as tangential force fluctuations, positive-torque fluctuations, and roll-over torque fluctuations. The accuracy of the sensors of the low-speed uniformity tester does not meet the test requirements of these parameters, and does not fully reflect the uniformity characteristics of the tires; these parameters can be used to characterize the uniformity of the tires at high speed. In a tire uniformity test, there are many factors that affect the unevenness of the tire, such as speed, load, and pressure, which should be examined in order to study the change of forces and moments when the tire is rolling at high speed. In this paper, a high-speed uniformity tester is used to measure the change of forces and moments of a tire under different speeds and different pressures, and the influence of speed and pressure on tire uniformity parameters is analyzed. Additionally, the relationship between tire uniformity parameters and vehicle performance is briefly described, providing a reference for improving the production level of tires and vehicle design.
Tire Uniformity Definition
A tire is an annular elastic body made of various materials such as rubber, steel cord, and polyester. Its quality, size, and rigidity are uneven during the structural design and production process. Tire uniformity testing means that the tire is inspected for uneven size, mass, and force under certain conditions of pressure, load, and speed [4]. Tire uniformity measurement involves the rotation of the tire by applying a load on a rotating drum that is in contact with the tire. Force sensors mounted beside the tire shaft measures the fluctuation of forces and moments in the circumferential direction. The unevenness of the tire mainly includes uneven mass, uneven size, and uneven rigidity [5]. Figure 1 shows a schematic of the unevenness of a tire. When a vehicle is running, uneven mass will result in dynamic and static imbalances, causing the tires to generate undulating forces and moments when rotating. The magnitudes of the forces are related to mass imbalance, shape imbalance, and stiffness imbalance. Unbalanced forces in the plane of rotation produce radial, tangential, and lateral excitation forces. Uneven and asymmetrical mass distribution about the rotating shaft causes a coupling imbalance, which produces a rotational moment on the wheel, which is expressed as a returning moment and a turning moment of the frequency change when the wheel rotates [6]. For uneven tire mass, a weight can be added to the rim to reduce the impact on vehicle performance.
A tire is an elastomer that can be characterized by a radial spring in the circumferential direction, and unevenness in stiffness may occur in the circumferential direction, mainly manifested as fluctuations in circumferential direction force, taper effect force, and angle effect force [7]. Uneven size will cause radial runout and lateral runout, and the tire will be polygonal in the process of rolling, instead of circular. It can affect the vibration or sway of the vehicle when riding.
Definition of Tire Uniformity Parameters
(1) The radial force (RF) is the force perpendicular to the spindle of the high-speed uniformity test machine, indicating the amount of load applied to the tire.
(2) The lateral force (LF) is the force parallel to the spindle of the high-speed uniformity test machine, reflecting the vehicle's handling performance.
(3) Tangential force (TF) is the driving force of tire and high-speed uniformity testing machine, reflecting the driving performance of the vehicle.
(4) Radial force fluctuation (RFV) refers to the fluctuation of radial force during one or more cycles of positive or negative rotation of a tire under a certain load, pressure, and speed, showing the upper and lower runout of the tire.
(5) Lateral force fluctuation (LFV) refers to the fluctuation of the inner side of the tire during one or more cycles of positive or negative rotation of the tire under a certain load, pressure, and speed, which is manifested as the left and right deviation of the car.
(6) Tangential force fluctuation (TFV) refers to the fluctuation of the driving direction of the tire during one or more cycles of positive or negative rotation under a certain load, pressure, and speed.
(7) Radial force 1st-10th harmonic (H1RFV-H10RFV) The relationship between the radial force of the tire and the angle of rotation of the tire obtained by the force fluctuation test is a resonance curve. The tire rotates forward or reverses one or more cycles. Force fluctuations can be decomposed into 1st to 10th harmonics by Fourier transform, where the primary component of the original waveform is called the fundamental or first harmonic.
(8) Lateral force 1st-10th harmonic (H1LFV-H10LFV) The relationship between the tire lateral force and the tire rotation angle obtained by the force fluctuation test is a resonance curve, and the tire is rotated forward or reversed one or more cycles. Force fluctuations can be decomposed into 1st to 10th harmonics by Fourier transform, where the primary component of the original waveform is called the fundamental or first harmonic.
(9) Tangential force 1st-10th harmonic (H1TFV-H10TFV) The relationship between the tangential force of the tire and the angle of rotation of the tire obtained by the force fluctuation test is a resonance curve. The tire is rotated forward or reversed for one or more cycles. Force fluctuations can be decomposed into 1st to 10th harmonics by Fourier transform, where the primary component of the original waveform is called the fundamental or first harmonic.
(10) Lateral force offset (LSFT) is the average value of the force integral on the inside of one or more cycles of forward or reverse rotation, reflecting the quality of the lateral uniformity of the tire. (11) The taper effect (CON) does not change the lateral force offset of the sign due to the change in the direction of rotation of the tire, reflecting the force on the shoulders of the tire. (12) The angular effect (PLY) changes the lateral force offset of the sign as the direction of rotation of the tire changes, reflecting the structural design of the belt.
(13) Torque fluctuation (SATV) is the fluctuation of the restoring torque to restore the steering wheel to the straight driving position relative to the reference value. (14) Flip torque fluctuation (OTTV) is the fluctuation of the rolling moment in the tire straight line direction from the reference value. Figure 2 shows the partial uniformity parameter.
Vehicles 2020, 2, FOR PEER REVIEW 4 of 14 (14) Flip torque fluctuation (OTTV) is the fluctuation of the rolling moment in the tire straight line direction from the reference value. Figure 2 shows the partial uniformity parameter.
Working Principle of the High-Speed Uniformity Testing Machine
As shown in Figure 3, the uniformity measurement of the tire is performed on a high-speed test machine. An AC servo motor drives the drum to rotate and applies a certain load to the tire to cause the tire to roll; a DC motor keeps the rotation speed of the drum within specified limits. Absolute encoders are mounted on the drum spindle and the tire shaft [8]. When the tire spindle rotates for one cycle, the absolute encoder uniformly distributes 1024 pulses; each sensor measures the fluctuation of force and torque. A computer records the output value of each sensor, until the data is collected. Using the drum surface of the high-speed uniformity test machine to simulate the running condition of the road surface, the drum is used to drive the rotation of the tire; the tire shaft fixes the tire and fixes the rim with a flange to avoid the tire oscillating during the rolling process. Under standard test conditions, a certain force is applied to the test tire by a hydraulic servo-loading system according to a specified load, and the motor drives the drum to drive the tire to roll at a prescribed speed. The sensor detection system performs high-speed uniformity testing of the tire, measurement of parameters, and transmittal of the results of the test to the computer, as shown in Figure 4.
Working Principle of the High-Speed Uniformity Testing Machine
As shown in Figure 3, the uniformity measurement of the tire is performed on a high-speed test machine. An AC servo motor drives the drum to rotate and applies a certain load to the tire to cause the tire to roll; a DC motor keeps the rotation speed of the drum within specified limits. Absolute encoders are mounted on the drum spindle and the tire shaft [8]. When the tire spindle rotates for one cycle, the absolute encoder uniformly distributes 1024 pulses; each sensor measures the fluctuation of force and torque. A computer records the output value of each sensor, until the data is collected.
Vehicles 2020, 2, FOR PEER REVIEW 4 of 14 (14) Flip torque fluctuation (OTTV) is the fluctuation of the rolling moment in the tire straight line direction from the reference value. Figure 2 shows the partial uniformity parameter.
Working Principle of the High-Speed Uniformity Testing Machine
As shown in Figure 3, the uniformity measurement of the tire is performed on a high-speed test machine. An AC servo motor drives the drum to rotate and applies a certain load to the tire to cause the tire to roll; a DC motor keeps the rotation speed of the drum within specified limits. Absolute encoders are mounted on the drum spindle and the tire shaft [8]. When the tire spindle rotates for one cycle, the absolute encoder uniformly distributes 1024 pulses; each sensor measures the fluctuation of force and torque. A computer records the output value of each sensor, until the data is collected. Using the drum surface of the high-speed uniformity test machine to simulate the running condition of the road surface, the drum is used to drive the rotation of the tire; the tire shaft fixes the tire and fixes the rim with a flange to avoid the tire oscillating during the rolling process. Under standard test conditions, a certain force is applied to the test tire by a hydraulic servo-loading system according to a specified load, and the motor drives the drum to drive the tire to roll at a prescribed speed. The sensor detection system performs high-speed uniformity testing of the tire, measurement of parameters, and transmittal of the results of the test to the computer, as shown in Figure 4. Using the drum surface of the high-speed uniformity test machine to simulate the running condition of the road surface, the drum is used to drive the rotation of the tire; the tire shaft fixes the tire and fixes the rim with a flange to avoid the tire oscillating during the rolling process. Under standard test conditions, a certain force is applied to the test tire by a hydraulic servo-loading system according to a specified load, and the motor drives the drum to drive the tire to roll at a prescribed speed. The sensor detection system performs high-speed uniformity testing of the tire, measurement of parameters, and transmittal of the results of the test to the computer, as shown in Figure 4.
condition of the road surface, the drum is used to drive the rotation of the tire; the tire shaft fixes the tire and fixes the rim with a flange to avoid the tire oscillating during the rolling process. Under standard test conditions, a certain force is applied to the test tire by a hydraulic servo-loading system according to a specified load, and the motor drives the drum to drive the tire to roll at a prescribed speed. The sensor detection system performs high-speed uniformity testing of the tire, measurement of parameters, and transmittal of the results of the test to the computer, as shown in Figure 4.
Mathematical Model of Uniformity Testing Machine
The measuring principle of radial force and lateral force is shown in Figure 5. Two vector force sensors are at points A and B [9]. The sensor acts at the support point of the two ends of the load axle; the forces of the sensor and the tire are as shown in the figure. Point O is the midpoint of the axis of the load wheel, and point C is the center of the contact surface between the tire and the load wheel.
Mathematical Model of Uniformity Testing Machine
The measuring principle of radial force and lateral force is shown in Figure 5. Two vector force sensors are at points A and B [9]. The sensor acts at the support point of the two ends of the load axle; the forces of the sensor and the tire are as shown in the figure. Point O is the midpoint of the axis of the load wheel, and point C is the center of the contact surface between the tire and the load wheel. (1) Figure 5 illustrates torque balance analysis that defines the x and y forces between the tire and the load wheel as and , that is, the radial and lateral forces measured for uniformity. During the test, the upper and lower force sensors are always in a static state, and the center position C of the tire and the load wheel contact remains unchanged. The three points A, B, and C are used for the torque balance: From ∑ = 0: The following can be derived from the above equations: From ∑ = 0: From the above equations, we can derive: According to Figure 5, we have: AO + BO = AB (1) Figure 5 illustrates torque balance analysis that defines the x and y forces between the tire and the load wheel as F r and F l , that is, the radial and lateral forces measured for uniformity. During the test, the upper and lower force sensors are always in a static state, and the center position C of the tire and the load wheel contact remains unchanged. The three points A, B, and C are used for the torque balance: From M A = 0: The following can be derived from the above equations: Vehicles 2020, 2 From the above equations, we can derive: where F r is radial force; F l is lateral force; F t is tangential force; F 0 is tire centrifugal force; F ur , F ul , F dr , F dl are the support forces of the upper and lower load cells in the x-direction and the y-direction, that is, the values measured by the sensors at the load; φ is the angle between the centrifugal force F 0 and the x-axis on the tire plane; and R is the load wheel radius. It can be derived from Equation (7) that the radial force of the tire is equal to the sum of the x-direction force of the upper and lower load cells and the component force of the centrifugal force in the x-direction, and that the lateral force is equal to the sum of the y-direction forces of the upper and lower sensors.
During the high-speed rotation of tires, the centrifugal force generated by the eccentric mass is one of the main sources of excitation. Regarding the drum and the tire as a vibration system, the mass of the drum is M, the eccentric mass of the tire is m, the eccentric distance is e, the angular velocity of the tire is ω, the stiffness of the two supporting ends of the sensor is k/2, and the damping coefficient is c. Let x be the distance of the tire drum system from the equilibrium position on the x-coordinate axis, the displacement of the eccentric mass is The differential equation of motion of the system on the x-coordinate axis is ..
Organize available M ..
The high-speed rotating eccentric mass tire can be regarded as the motion under the action of simple harmonic excitation, where is the amplitude of the exciting force.
The steady-state response of forced vibration of the system under harmonic excitation where B is the amplitude of steady-state vibration The phase difference between the displacement x of the vibrating object and the excitation force: Frequency ratio λ = ω ω n , relative damping coefficient ξ = c 2ω nM , and natural frequency of the system without damping ω n = k M . The amplitude B is directly proportional to the system imbalance. In order to reduce vibration, high-speed rotating parts must be tested for dynamic balance to make the mass as uniform as possible.
When the unbalance of the system cannot be compensated, if the frequency of the excitation force, that is λ = 1, the angular velocity of the motor rotor is close to the natural frequency of the system, that is, the system will have the strongest vibration, that is, resonance. At this time, the motor speed is called critical speed, and output power is called limit power. The resonance phenomenon has a great impact on the performance and life of mechanical equipment and structures, and even directly causes damage to components or the entire mechanical equipment. When the system is in resonance, the maximum amplitude is inversely proportional to the damping. This provides a measure of vibration reduction in theory. In order to avoid the occurrence of resonance, in the design, the power of the motor should be avoided as the limit power, or the motor speed should be kept away from the critical speed of the system.
Uniformity Parameter Signal Acquisition
There is a plurality of load cells on the main shaft of the high-speed uniformity tire tester. When the tire rotates, the force and torque output signals of the radial, lateral, and tangential forces of the tire pass through the signal amplifier, which performs signal separation and conversion. The amplified force and torque ripple signals are extracted by the low pass filter to extract the required frequency components, which are sent to the AD converter and received by the data collector, and then Fourier-transformed by the data collector. After correction of the low pass filter and calculation of the force and phase angle at various frequencies and the force fluctuation of the tire rotation, the data collector transmits the results to the computer in an appropriate form [10]. Figure 6 shows a schematic of the data collection for uniformity parameters.
Vehicles 2020, 2, FOR PEER REVIEW 7 of 14 frequency components, which are sent to the AD converter and received by the data collector, and then Fourier-transformed by the data collector. After correction of the low pass filter and calculation of the force and phase angle at various frequencies and the force fluctuation of the tire rotation, the data collector transmits the results to the computer in an appropriate form [10]. Figure 6 shows a schematic of the data collection for uniformity parameters.
Method for Calculating Uniformity Parameters
The tire is filled with a certain air pressure and, under certain load conditions, the tire is rotated at a certain speed by the drum spindle. In one cycle, the absolute encoder sends 1024 pulses; each time the encoder sends a pulse, the computer records the output of each sensor, continuing until the required data is collected [11]. The waveform of the circumferential force fluctuation is obtained by filtering, and the amplitude, phase, and frequency of the different sine or cosine wave components of the signal are calculated by using the measured original signal according to the Fourier transform. In physics, the simplest vibrational wave is a simple harmonic represented by a sinusoidal function, that is sin(2 + ), where is the amplitude, is the frequency, and is the phase angle. More elaborate waves are obtained when several simple waves are added together. Similarly, a nonsimple waveform can be transformed into a sum of simple wave of different orders by Fourier transform [12].
An arbitrary wave ( ), whose period is and whose angular frequency is = 2 , can be decomposed into the following: ( ) = 2 + cos( + ) + cos( + ) + ⋯ which is equivalent to: From the orthogonal relationship of the sine and cosine functions, the Fourier coefficients can be obtained:
Method for Calculating Uniformity Parameters
The tire is filled with a certain air pressure and, under certain load conditions, the tire is rotated at a certain speed by the drum spindle. In one cycle, the absolute encoder sends 1024 pulses; each time the encoder sends a pulse, the computer records the output of each sensor, continuing until the required data is collected [11]. The waveform of the circumferential force fluctuation is obtained by filtering, and the amplitude, phase, and frequency of the different sine or cosine wave components of the signal are calculated by using the measured original signal according to the Fourier transform. In physics, the simplest vibrational wave is a simple harmonic represented by a sinusoidal function, that is A sin(2π f t + ϕ), where A is the amplitude, f is the frequency, and ϕ is the phase angle. More elaborate waves are obtained when several simple waves are added together. Similarly, a non-simple waveform can be transformed into a sum of simple wave of different orders by Fourier transform [12].
An arbitrary wave f (t), whose period is T and whose angular frequency is ω = 2π f , can be decomposed into the following: which is equivalent to: From the orthogonal relationship of the sine and cosine functions, the Fourier coefficients can be obtained: Combine the above formulas: which is equivalent to: where A 0 = a 0, A n = a 2 n + b 2 n , ϕ n = −arctan b n a n .
A 0 2 is the DC component of f (t), f 0 = 1 T is the baseband frequency, A n cos(nωt + ϕ n ) is the nth harmonic of f (t), A n is the amplitude, and ϕ n is the initial phase angle of the nth harmonic. When the signal is analyzed and processed, the signal can be decomposed into superposed harmonics according to the Fourier transform, and the amplitude and phase angle of each harmonic are respectively obtained. According to the transformed result, the harmonics of the positive and negative directions of the radial force, the lateral force, and the traction force can be known.
When calculating the lateral force offset LSTF, we assume that the lateral force forward offset is LSTF cw , and the lateral force reversal offset is LSTF ccw ; thus: According to the above formulas, the numerical values of the tire uniformity parameters can be obtained, so that the quality of the tire can be evaluated, the amplitude of the force value fluctuation can be determined, and the tire manufacturing process can be improved, providing a basis for tire companies to improve tire uniformity and improve the performance of vehicles in the automotive design department [13].
High-Speed Uniformity Test and Analysis of Tires
Before the tire uniformity test, it is necessary to define the load, pressure, and speed of the tire, and apply a certain load to the tire according to the use of the tire and the production process requirements. A test was undertaken to simulate the load of a vehicle when measuring the uniformity of a tire with a mass of 75 kg; the tire size used was 205/55R16. The tire load was set to 4221 N, 70% of the specified maximum load, which is expressed as the average value of the load received by the tire during one rotation. According to the demands of use, the tire is filled with a certain inflation pressure; the pressures used in this paper were 0.16 MPa, 0.20 MPa, and 0.24 MPa, expressed as the internal pressure of the tire that is fed back through the barometer during the test. According to the corresponding control system of the high-speed uniformity testing machine, the inflation pressure of the tire remains unchanged during the test; this paper studies the change of the force and moment of the tire when it is rolling at high speed, so the speed is set to vary from 120 km/h to 240 km/h in increments of 20 km/h.
Relationship between Velocity and Uniformity Parameters
A tire was tested at varying speeds by the high-speed uniformity tester, which recorded relevant characteristic parameters, as shown in Table 1. The load was 4221 N, the pressure was 0.20 MPa, and the speeds were as shown in the table. Based on the parameters of Table 1, the relationship between velocity and radial force was plotted. It can be seen from Figure 7 that, as the speed increased, the radial force fluctuation and the radial force first harmonic gradually increased. As the speed increased from 120 km/h to 240 km/h, the RFV increased from 213 N to 803 N, and the H1RFV increased from 152.9 N to 553.4 N; that is, the growth rate of RFV was greater than the growth rate of its first harmonic. When the vehicle is driving at a high speed, the vibration of the car is enhanced, noise is increased, and it is easy for the car to jump up and down, thereby affecting ride comfort and causing early damage of the automobile parts. According to previous studies by scholars [14], the radial force fluctuation of the tire and the first harmonic of the radial force are inherent characteristics of the tire under a certain inflation pressure. The load change has no effect on it. When the vehicle fluctuates up and down, it is necessary to replace the tires with better quality tires in order to ensure the comfort and safety of the ride. Based on the parameters of Table 1, the relationship between velocity and radial force was plotted. It can be seen from Figure 7 that, as the speed increased, the radial force fluctuation and the radial force first harmonic gradually increased. As the speed increased from 120 km/h to 240 km/h, the RFV increased from 213 N to 803 N, and the H1RFV increased from 152.9 N to 553.4 N; that is, the growth rate of RFV was greater than the growth rate of its first harmonic. When the vehicle is driving at a high speed, the vibration of the car is enhanced, noise is increased, and it is easy for the car to jump up and down, thereby affecting ride comfort and causing early damage of the automobile parts. According to previous studies by scholars [14], the radial force fluctuation of the tire and the first harmonic of the radial force are inherent characteristics of the tire under a certain inflation pressure. The load change has no effect on it. When the vehicle fluctuates up and down, it is necessary to replace the tires with better quality tires in order to ensure the comfort and safety of the ride.
Speed vs. LFV and H1LFV
It can be seen from Figure 8 that there is no obvious relationship between velocity and either the fluctuation of the lateral force or its first harmonic at speeds less than 220 km/h. According to the test data, at the speed of 220 km/h, a standing wave phenomenon had occurred in the tire. At speeds greater than 220 km/h, the lateral force fluctuation gradually increased with speed. The steering wheel of the car at high speed is easy to swing, which should enhance stability during driving and help avoid traffic accidents.
Speed vs. LFV and H1LFV
It can be seen from Figure 8 that there is no obvious relationship between velocity and either the fluctuation of the lateral force or its first harmonic at speeds less than 220 km/h. According to the test data, at the speed of 220 km/h, a standing wave phenomenon had occurred in the tire. At speeds greater than 220 km/h, the lateral force fluctuation gradually increased with speed. The steering wheel of the car at high speed is easy to swing, which should enhance stability during driving and help avoid traffic accidents.
Speed vs. LFV and H1LFV
It can be seen from Figure 8 that there is no obvious relationship between velocity and either the fluctuation of the lateral force or its first harmonic at speeds less than 220 km/h. According to the test data, at the speed of 220 km/h, a standing wave phenomenon had occurred in the tire. At speeds greater than 220 km/h, the lateral force fluctuation gradually increased with speed. The steering wheel of the car at high speed is easy to swing, which should enhance stability during driving and help avoid traffic accidents.
Speed vs. SATV and H1SATV
It can be seen from Figure 9 that, when the speed was at the interval 120-190 km/h, the fluctuation of the positive moment increased gradually with the increase of the speed, but when the speed exceeded 190 km/h, the fluctuation of the positive moment suddenly dropped. When the speed of tire rolling does not exceed 190 km/h, the steering stability of the car gradually decreases with the increase of speed, and the steering direction of the car is limited [15]. When the car is driving at a high speed, the tire is easily swayed, and the steering stability and high-speed driving safety are deteriorated. In addition, the safety of high-speed driving is worse, and the wear resistance of the tire is degraded. It can be seen from Figure 9 that, when the speed was at the interval 120-190 km/h, the fluctuation of the positive moment increased gradually with the increase of the speed, but when the speed exceeded 190 km/h, the fluctuation of the positive moment suddenly dropped. When the speed of tire rolling does not exceed 190 km/h, the steering stability of the car gradually decreases with the increase of speed, and the steering direction of the car is limited [15]. When the car is driving at a high speed, the tire is easily swayed, and the steering stability and high-speed driving safety are deteriorated. In addition, the safety of high-speed driving is worse, and the wear resistance of the tire is degraded.
Speed vs. TFV and H1TFV
It can be seen from Figure 10 that the relationship between velocity and tangential force fluctuation is approximately linear. When the tire was rolling at high speed, the tangential force increased rapidly; as the speed increased from 120 km/h to 240 km/h, the tangential force fluctuation increased from 208 N to 935 N. As the tangential force of the tire increases, the tangential deformation increases, and the frictional effect of the contact surface increases, thereby suppressing the tendency of slipping. The car's braking, power, and handling can be expected to function properly.
Speed vs. TFV and H1TFV
It can be seen from Figure 10 that the relationship between velocity and tangential force fluctuation is approximately linear. When the tire was rolling at high speed, the tangential force increased rapidly; as the speed increased from 120 km/h to 240 km/h, the tangential force fluctuation increased from 208 N to 935 N. As the tangential force of the tire increases, the tangential deformation increases, and the frictional effect of the contact surface increases, thereby suppressing the tendency of slipping. The car's braking, power, and handling can be expected to function properly.
It can be seen from Figure 10 that the relationship between velocity and tangential force fluctuation is approximately linear. When the tire was rolling at high speed, the tangential force increased rapidly; as the speed increased from 120 km/h to 240 km/h, the tangential force fluctuation increased from 208 N to 935 N. As the tangential force of the tire increases, the tangential deformation increases, and the frictional effect of the contact surface increases, thereby suppressing the tendency of slipping. The car's braking, power, and handling can be expected to function properly. It can be seen from Figure 11 that the fluctuation of the turning moment gradually increased with the increase of the speed. At a high speed, the vehicle is prone to swinging left and right. Due to the uneven radial stiffness of the tire, the dimensional deviation of the hub causes the wheel to roll at the axle. Vertical excitation is generated, causing the steering wheel to reduce the interest of driving.
Speed vs. OTTV and H1OTTV
It can be seen from Figure 11 that the fluctuation of the turning moment gradually increased with the increase of the speed. At a high speed, the vehicle is prone to swinging left and right. Due to the uneven radial stiffness of the tire, the dimensional deviation of the hub causes the wheel to roll at the axle. Vertical excitation is generated, causing the steering wheel to reduce the interest of driving.
Relationship between Pressure and Uniformity Parameters
Testing by high-speed uniformity tester and recording relevant characteristic parameters, as shown in Table 2, the load is 4221 N, the speed is 120 km/h, and the pressures are as shown in the table. Table 2. Relationship between tire pressure and tire uniformity parameters. It can be seen from Figure 12 that, as the pressure increased, the radial force fluctuation of the tire reached a minimum at 0.19 MPa, and the first harmonic of the radial force gradually increased. When the pressure increased from 0.16 MPa to 0.24 MPa, the radial force fluctuations increased from 217 N to 225 N; that is, the overall trend is rising. As the inflation pressure increased, the RFV and LFV of the tire gradually increased [16]. There are many reasons for fluctuations in RFV, such as variations in tread size of semi-finished parts, uneven density of carcass and belt cords, and dimensional variations of extruded parts such as sidewalls and apex-or, to cite other examples, the distribution of the joint of the forming process is offset, the joint quantity is too large or too small, or the length of the cord between the two beads is different, the blank of the vulcanization process is not centered with the vulcanization mold, the degree of vulcanization in the circumferential direction of the tire is not uniform, or the fillet storage deformation [17].
Relationship between Pressure and Uniformity Parameters
Testing by high-speed uniformity tester and recording relevant characteristic parameters, as shown in Table 2, the load is 4221 N, the speed is 120 km/h, and the pressures are as shown in the table. It can be seen from Figure 12 that, as the pressure increased, the radial force fluctuation of the tire reached a minimum at 0.19 MPa, and the first harmonic of the radial force gradually increased. When the pressure increased from 0.16 MPa to 0.24 MPa, the radial force fluctuations increased from 217 N to 225 N; that is, the overall trend is rising. As the inflation pressure increased, the RFV and LFV of the tire gradually increased [16]. There are many reasons for fluctuations in RFV, such as variations in tread size of semi-finished parts, uneven density of carcass and belt cords, and dimensional variations of extruded parts such as sidewalls and apex-or, to cite other examples, the distribution of the joint of the forming process is offset, the joint quantity is too large or too small, or the length of the cord between the two beads is different, the blank of the vulcanization process is not centered with the vulcanization mold, the degree of vulcanization in the circumferential direction of the tire is not uniform, or the fillet storage deformation [17].
tire reached a minimum at 0.19 MPa, and the first harmonic of the radial force gradually increased. When the pressure increased from 0.16 MPa to 0.24 MPa, the radial force fluctuations increased from 217 N to 225 N; that is, the overall trend is rising. As the inflation pressure increased, the RFV and LFV of the tire gradually increased [16]. There are many reasons for fluctuations in RFV, such as variations in tread size of semi-finished parts, uneven density of carcass and belt cords, and dimensional variations of extruded parts such as sidewalls and apex-or, to cite other examples, the distribution of the joint of the forming process is offset, the joint quantity is too large or too small, or the length of the cord between the two beads is different, the blank of the vulcanization process is not centered with the vulcanization mold, the degree of vulcanization in the circumferential direction of the tire is not uniform, or the fillet storage deformation [17].
(a) (b) Figure 12. Relationship between pressure and radial force fluctuation.
Pressure vs. LFV and H1LFV
It can be seen from Figure
Pressure vs. LFV and H1LFV
It can be seen from Figure 13 that the LFV and H1LFV of the tire decreased in an approximately linear relationship with the increase of the pressure. The slope of the LFV curve in the figure is about −162.5. The amount of lateral deformation is reduced, so the lateral force fluctuation of the tire decreases. Some of the factors that cause the tire to produce LFV are as follows: belt layer angle and width variation or serpentine distortion of the semi-finished parts, apex height, unevenness in thickness, or thickness variation of the belt layer. Alternatively, the carcass ply of the forming process is reversely biased, snake twist when the belt is attached, or the belt cushion is conformed to the hemiplegia or dimensional variation during the forming-or the deformation of the tire in the vulcanization process leads to irregular vulcanization and large temperature deviation between the vulcanization molds, resulting in inconsistent degree of vulcanization of the upper and lower parts of the tire, insufficient processing precision of the vulcanization mold, and/or asymmetrical upper and lower dimensions [18].
Vehicles 2020, 2, FOR PEER REVIEW 12 of 14 −162.5. The amount of lateral deformation is reduced, so the lateral force fluctuation of the tire decreases. Some of the factors that cause the tire to produce LFV are as follows: belt layer angle and width variation or serpentine distortion of the semi-finished parts, apex height, unevenness in thickness, or thickness variation of the belt layer. Alternatively, the carcass ply of the forming process is reversely biased, snake twist when the belt is attached, or the belt cushion is conformed to the hemiplegia or dimensional variation during the forming-or the deformation of the tire in the vulcanization process leads to irregular vulcanization and large temperature deviation between the vulcanization molds, resulting in inconsistent degree of vulcanization of the upper and lower parts of the tire, insufficient processing precision of the vulcanization mold, and/or asymmetrical upper and lower dimensions [18].
(a) (b) Figure 13. Relationship between pressure and lateral force fluctuation.
Pressure vs. SATV and H1SATV
It can be seen from Figure 14 that, with the increase of pressure, the fluctuation of the positive moment gradually decreased. At pressures less than 0.22 MPa, the rate of fluctuation of the returning moment was faster. At pressures greater than 0.22 MPa, the fluctuation of the positive moment tended to gradually smooth. The trend of the first harmonic was basically consistent with the fluctuation of the positive torque. When the car is driven under low tire pressure, the steering stability of the car is deteriorated. Under standard tire pressure, the steering stability and steering sensitivity of the car can be expected to be satisfactory. Drivers should pay attention to tire pressure to ensure that the car performs properly.
Pressure vs. SATV and H1SATV
It can be seen from Figure 14 that, with the increase of pressure, the fluctuation of the positive moment gradually decreased. At pressures less than 0.22 MPa, the rate of fluctuation of the returning moment was faster. At pressures greater than 0.22 MPa, the fluctuation of the positive moment tended to gradually smooth. The trend of the first harmonic was basically consistent with the fluctuation of the positive torque. When the car is driven under low tire pressure, the steering stability of the car is deteriorated. Under standard tire pressure, the steering stability and steering sensitivity of the car can be expected to be satisfactory. Drivers should pay attention to tire pressure to ensure that the car performs properly. moment was faster. At pressures greater than 0.22 MPa, the fluctuation of the positive moment tended to gradually smooth. The trend of the first harmonic was basically consistent with the fluctuation of the positive torque. When the car is driven under low tire pressure, the steering stability of the car is deteriorated. Under standard tire pressure, the steering stability and steering sensitivity of the car can be expected to be satisfactory. Drivers should pay attention to tire pressure to ensure that the car performs properly.
Pressure vs. TFV and H1TFV
It can be seen from Figure 15 that, under certain speed conditions, the tangential force fluctuation of the tire increased with the increase of the pressure, which affects the ride comfort of the car.
Pressure vs. TFV and H1TFV
It can be seen from Figure 15 that, under certain speed conditions, the tangential force fluctuation of the tire increased with the increase of the pressure, which affects the ride comfort of the car.
Pressure vs. OTTV and H1OTTV
As shown in Figure 16, the tire's turning moment fluctuation gradually decreased with the increase of the pressure, which is caused by the uneven radial stiffness of the tire. At a certain driving speed, increasing the pressure can be effective. Reduce the left and right swing of the vehicle and improve the comfort of the ride.
Conclusions
Through the high-speed uniformity test of different speeds and different pressures, the uniformity parameter values of a tire under different working conditions were obtained, Furthermore, the influences of speed and pressure on the uniformity parameters of tires and the relationship between uniformity parameters and vehicle performance were obtained. According to these test results, the influence of speed and pressure on tire uniformity parameters can be known: (1) With the increase of speed, the RFV, LFV, TFV, and OTTV of the tire gradually increase, and the car becomes more prone to up-and-down and left-and-right swing; the SATV of the tire increases with the increase of speed within a certain speed range, but, once the speed exceeds a certain value,
Pressure vs. OTTV and H1OTTV
As shown in Figure 16, the tire's turning moment fluctuation gradually decreased with the increase of the pressure, which is caused by the uneven radial stiffness of the tire. At a certain driving speed, increasing the pressure can be effective. Reduce the left and right swing of the vehicle and improve the comfort of the ride.
Pressure vs. OTTV and H1OTTV
As shown in Figure 16, the tire's turning moment fluctuation gradually decreased with the increase of the pressure, which is caused by the uneven radial stiffness of the tire. At a certain driving speed, increasing the pressure can be effective. Reduce the left and right swing of the vehicle and improve the comfort of the ride.
Conclusions
Through the high-speed uniformity test of different speeds and different pressures, the uniformity parameter values of a tire under different working conditions were obtained, Furthermore, the influences of speed and pressure on the uniformity parameters of tires and the relationship between uniformity parameters and vehicle performance were obtained. According to these test results, the influence of speed and pressure on tire uniformity parameters can be known: (1) With the increase of speed, the RFV, LFV, TFV, and OTTV of the tire gradually increase, and the car becomes more prone to up-and-down and left-and-right swing; the SATV of the tire increases with the increase of speed within a certain speed range, but, once the speed exceeds a certain value,
Conclusions
Through the high-speed uniformity test of different speeds and different pressures, the uniformity parameter values of a tire under different working conditions were obtained, Furthermore, the influences of speed and pressure on the uniformity parameters of tires and the relationship between uniformity Vehicles 2020, 2 572 parameters and vehicle performance were obtained. According to these test results, the influence of speed and pressure on tire uniformity parameters can be known: (1) With the increase of speed, the RFV, LFV, TFV, and OTTV of the tire gradually increase, and the car becomes more prone to up-and-down and left-and-right swing; the SATV of the tire increases with the increase of speed within a certain speed range, but, once the speed exceeds a certain value, the value of SATV gradually decreases with further increases of speed, and the car is prone to deviation in a certain speed range.
(2) With the increase of pressure, the RFV and TFV of the tire gradually become larger; the LFV, SATV, and OTTV of the tire decrease with the increase of pressure.
(3) Compared with the low-speed uniformity test, the high-speed uniformity test can detect the tangential force fluctuation of the tire, the positive torque fluctuation, and the turning torque fluctuation, which can fully characterize the uniformity of the tire.
The tire uniformity parameters directly affect the driving safety of the car and can be used to improve the uniformity of the tires, especially radial tires under high-speed driving conditions, which is of great significance for tire and automobile design companies to improve the performance of vehicles. These results may help a tire designer to select the most suitable solution when designing the structure of a tire, and at the same time help tire testers to set more reasonable test conditions, so that the tested tire can perform more accurately under the most reasonable conditions, which will help the improvement of vehicle performance.
Conflicts of Interest:
This manuscript did not lead to any conflicts of interest regarding the publication. | 12,380.2 | 2020-09-08T00:00:00.000 | [
"Engineering"
] |
Integrated Metabolomic and Transcriptomic Analysis of Puerarin Biosynthesis in Pueraria montana var. thomsonii at Different Growth Stages
Puerarin, a class of isoflavonoid compounds concentrated in the roots of Puerarias, has antipyretic, sedative, and coronary blood-flow-increasing properties. Although the biosynthetic pathways of puerarin have been investigated by previous researchers, studies focusing on the influence of different growth stages on the accumulation of metabolites in the puerarin pathway are not detailed, and it is still controversial at the last step of the 8-C-glycosylation reaction. In this study, we conducted a comprehensive analysis of the metabolomic and transcriptomic changes in Pueraria montana var. thomsonii during two growing years, focusing on the vigorous growth and dormant stages, to elucidate the underlying mechanisms governing the changes in metabolite and gene expression within the puerarin biosynthesis pathway. In a comparison of the two growth stages in the two groups, puerarin and daidzin, the main downstream metabolites in the puerarin biosynthesis pathway, were found to accumulate mainly during the vigorous growth stage. We also identified 67 common differentially expressed genes in this pathway based on gene expression differences at different growth stages. Furthermore, we identified four candidate 8-C-GT genes that potentially contribute to the conversion of daidzein into puerarin and eight candidate 7-O-GT genes that may be involved in the conversion of daidzein into daidzin. A co-expression network analysis of important UGTs and HIDs along with daidzein and puerarin was conducted. Overall, our study contributes to the knowledge of puerarin biosynthesis and offers information about the stage at which the 8-C-glycosylation reaction occurs in biosynthesis. These findings provide valuable insights into the cultivation and quality enhancement of Pueraria montana var. thomsonii.
Introduction
Pueraria montana var.thomsonii (Bentham) M. R. Almeida (hereafter referred to as P. montana var.thomsonii) is a perennial climbing vine of the Leguminosae family [1,2].The Pueraria genus comprises more than 20 species, mainly distributed in temperate and subtropical regions [2,3].P. montana var.thomsonii is a notable medicinal plant known for its pharmacological activities, including heart and brain protection, hepatoprotection, and hypotensive and hypoglycemic properties, as well as antitumor capabilities [4].It contains various chemical compounds, such as isoflavonoids, triterpenoids, alkaloids, saponins, coumarins, and polysaccharides [5].The primary medicinal compounds in the root tubers of P. montana var.thomsonii are flavonoids and isoflavones, including daidzein, genistein, daidzin (7-O-glycoside of daidzein), puerarin (8-C-glycoside of daidzein), etc. [6].In addition to being used as a medicine, P. montana var.thomsonii is also in great demand as a food and nutraceutical, and it is popular in Asian countries, especially Japan and Thailand [7].
Plant development is a complex process, and the metabolites accumulated in plants during different developmental stages vary greatly [25].However, most of the previous studies have focused on different tissues or different varieties to identify candidate genes from differentially expressed genes, and few studies have been conducted to investigate the transcriptomic and metabolomic profiles of P. montana var.thomsonii at different growth stages to probe the accumulation mechanisms of isoflavonoids in the puerarin biosynthesis pathways from the perspective of growth stages.
In this study, the differences in metabolites and genes in P. montana var.thomsonii root tubers during vigorous growth and dormant stages were investigated using ultraperformance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS) and RNA sequencing (RNA-Seq) analysis techniques.Root tubers from one-year-old and two-yearold P. montana var.thomsonii from two growth stages were compared separately to obtain differentially expressed genes and differentially accumulated metabolites associated with growth stages that were common to both the one-and two-year-old sample sets, thus excluding other interferences and obtaining more reliable data.A detailed analysis of the gene expression data and phylogenetic relationship allowed us to propose important UGTs, including several C-UGT candidates and O-UGT candidates, that are likely involved in the downstream biosynthesis and regulation of puerarin and daidzin in P. montana var.thomsonii.It also offers information about the stage at which the 8-C-glycosylation reaction occurs during biosynthesis.In conclusion, we systematically investigated the metabolite accumulation and molecular mechanisms of puerarin biosynthesis during different growth stages of the root tuber of P. montana var.thomsonii using metabolomics and transcriptomics approaches combined with co-expression analysis, which provided new insights into the regulation of isoflavonoids and established a foundation for quality enhancement of P. montana var.thomsonii.
Plant Materials
The samples used in this study were collected from healthy P.montana var.thomsonii plants growing in Zhongxiang City, Hubei Province, China (112 • 54 ′ 33 ′′ E, 31 • 21 ′ 12 ′′ N, with an altitude of 154.3 m).Fresh root tubers from two growth years at two growth stages were separately gathered.The basic information of the samples and the content of puerarin are shown in Table 1."GX" is defined as the vigorous growth stage (September) group, "GD" as the dormant stage (January) group, and the numbers following represent the growth years.For each growth stage at which the plant materials were sampled, three biological replicates with three technical repeats were performed, totaling 12 sets of samples.A total of 3 g of each set of samples was collected in EP tubes, frozen on dry ice, stored at −80 • C until RNA, and metabolites were extracted.The remaining samples were dried and broken into a powder that could pass through a No. 3 sieve, which was subsequently made into a test solution to determine the content of puerarin using high-performance liquid chromatography.
Processing of Samples and Procedures of HPLC
The chromatograph was a Shimadzu L-16, and the column was manufactured by Dalian Elite Analytical Instruments Co., Ltd., Dalian, China.Octadecylsilane-bonded silica gel was used as filler; methanol-water (25:75) was used as the mobile phase; the detection wavelength was 250 nm.
In detail, a sample powder amount of about 0.8 g was taken, weighed precisely, put in a stoppered conical flask, mixed with 50 mL of 30% ethanol, and heated under reflux for 30 min; after cooling, it was filtered, and the filtrate was used to make the test solution.
Puerarin standard was purchased from Lemeitan Pharmaceuticals, and 30% ethanol was added to make a solution containing 80 µg per 1 mL, which served as the control solution.
A volume of 10 µL of both the control solution and test solution was injected into the liquid chromatograph, and determination was carried out [26].
RNA Extraction and RNA Sequencing
Ethanol precipitation and CTAB-PBIOZOL were used for extraction, and the successfully extracted RNA was dissolved by adding 50 µL of DEPC-treated water.Subsequently, total RNA was identified and quantified using NanoDrop and Agilent 2100 Bioanalyzer (Thermo Fisher Scientific, Waltham, MA, USA).After constructing the mRNA library, sequencing was performed with the Illumina platform.The downstream data were filtered to obtain clean data, and sequence comparison was performed with the reference genome to obtain mapped data, which were subsequently used to perform structural-level analysis, such as variable splicing analysis, new gene discovery gene structure optimization, and expression-level analysis, such as differential expression analysis, functional annotation of differentially expressed genes, and functional enrichment, based on the expression of the genes in different samples or different groups of samples.Our study was based on the genomic data of P. montana var.thomsonii from the Guangxi Academy of Agricultural Sciences, China, as a reference (GenBank ID: GCA_019096045.1).
UPLC-MS/MS-Based Widely Targeted Metabolomic Analysis
The samples were placed in a lyophilizer (Scientz-100F, Scientz, Ningbo, China) for vacuum freeze drying and then ground for 1.5 min to powder using a grinder (MM 400, Retsch, Haan, German) at 30 Hz.Then, 50 mg of powder was weighed using an electronic balance (MS105DM), and 1200 µL of pre-cooled 70% methanol aqueous internal standardized extract at −20 • C was added.The extract was vortexed once every 30 min for 30 s for a total of 6 times and centrifuged at 12,000 rpm for 3 min.The supernatant was aspirated, and the sample was filtered through a microporous filter membrane (0.22 µm pore size) and preserved in the injection vial for UPLC-MS/MS analysis.
Liquid phase detection was performed on an Agilent SB-C18 1.8 µm, 2.1 mm × 100 mm column, with solvent A (pure water with 0.1% formic acid) and solvent B (acetonitrile with 0.1% formic acid).The flow rate was 0.35 mL/min, the column temperature was 40 • C, the injection volume was 2 µL, and the elution was based on a gradient.The mass spectrometry conditions were as follows: electrospray ionization (ESI) temperature, 500 • C; ion spray voltage (IS), 5500 V (positive ion mode)/−4500 V (negative ion mode).The ion source gas I (GSI), gas II (GSII), and curtain gas (CUR) were set to 50, 60, and 25 psi, respectively, and the collision-induced ionization parameter was set to high.QQQ scans were performed using multiple reaction monitoring (MRM) mode with the collision gas (nitrogen) set to medium.DP and CE of individual MRM ion pairs were accomplished by further de-clustering potential (DP) and collision energy (CE) optimization.A specific set of MRM ion pairs was monitored in each period based on the metabolites eluted within each period.
The quantification of metabolites was performed using triple quadrupole mass spectrometry.In MRM mode, the quadrupole first screens the precursor ions of the target substance and excludes ions corresponding to other molecular weight substances to preliminarily eliminate interference.Precursor ions are induced to ionize by colliding cells and then disintegrate to form many fragmented ions.Then, the fragment ions are filtered through a triple quadrupole to select the desired characteristic fragments, eliminate interference from non-target ions, and make quantification more accurate and repeatable.After obtaining spectral analysis data of metabolites from different samples, peak area integration was performed on all mass spectra peaks of the substance, and integration correction was performed on mass spectra peaks of the same metabolite in different samples [27].To compare the differences in the content of each metabolite in different samples among all detected metabolites, we corrected the mass spectrometry peaks detected for each metabolite in different samples based on the information on metabolite retention time and peak shape to ensure the accuracy of quantification and quantification.The relative content of metabolites in each sample was expressed as the integral of the chromatographic peak area.
Analysis of Differential Accumulation Metabolites and Differentially Expressed Genes
The variable importance in projection (VIP) value indicated the degree of influence of the intergroup difference of the corresponding metabolite in the classification discrimination of the samples in each group in the model, and the metabolites with VIP ≥ 1 were generally considered to have significant differences.Additionally, metabolites with a fold change (FC) greater than or equal to 2 or less than or equal to 0.5 between the control and experimental groups were considered significantly different.Overall, metabolites with VIP ≥ 1 and |log2 FC| ≥ 1 were defined as differentially accumulated metabolites (DAMs).
Moreover, differential expression analysis between sample groups was performed using DESeq2 to obtain the set of differentially expressed genes between the two comparison groups, and the false discovery rate (FDR) was obtained via multiple hypothesis testing corrected for the probability of hypothesis testing (p-value) using the Benjamini-Hochberg method.Differentially expressed genes (DEGs) were screened for |log2FC| ≥ 1 and FDR < 0.05.
Kyoto Encyclopedia of Genes and Genomes Annotation and Enrichment Analysis
Identified metabolites were annotated using the Kyoto Encyclopedia of Genes and Genomes (KEGG) compound database (http://www.kegg.jp/kegg/compound/,accessed on 1 June 2023), and annotated metabolites were then mapped to the KEGG Pathway database (http://www.kegg.jp/kegg/pathway.html,accessed on 1 June 2023).Pathways with significantly regulated metabolites were then subjected to metabolite set enrichment analysis (MSEA), and their significance was determined via hypergeometric test p-values.
Identification of Structural Genes in the Puerarin Biosynthesis Pathway
To identify all the structural genes involved in this pathway, known protein sequences were retrieved from the NCBI protein database (https://www.ncbi.nlm.nih.gov/protein/,accessed on 15 June 2023).These sequences were then searched on Interpro (https://www.ebi.ac.uk/interpro/search/sequence/, accessed on 15 June 2023) to obtain the Pfam numbers and target structural domains.Then, a biosequence analysis using profile hidden Markov models (HMMERs) was used to search sequence databases for homologous sequences and perform sequence alignments.Simultaneously, we employed local BLAST 2.14.0 software to identify the target structural genes within the P. montana var.thomsonii transcriptome of our current study.The results obtained from HMMER and BLAST were combined, and duplicate sequences were removed.Sequences with incomplete conserved structural domains were further screened using the CD-search (https://www.ncbi.nlm.nih.gov/Structure/cdd/wrpsb.cgi, accessed on 20 June 2023) online protein structure prediction website.Finally, manual screening was conducted to ensure the selection of relevant sequences.
Phylogenetic Analysis of UGTs in the Puerarin Biosynthesis Pathway
The glycosyltransferase (GT) protein superfamily contains numerous members, with family 1 (also known as ureido-diphosphate-glycosyltransferase UDP-glycosyltransferase (UGT)) being the largest and most closely related to plants [28].Due to the diversity and specificity of glycosyl acceptors and glycosyl donors, UGT can be divided into Oglycosyltransferases (OGTs) and C-glycosyltransferases (CGTs), which can be further subdivided according to the modification sites of glycosylation reactions.
To perform a phylogenetic analysis of the UGTs, protein sequences of previously characterized UGTs with known structures and functions from Pueraria montana var.lobata (referred to as P. montana var.lobata) or other species were downloaded from the NCBI protein database (https://www.ncbi.nlm.nih.gov/protein/,accessed on 25 June 2023).All UGT sequences were aligned using the MUSCLE algorithm, and a phylogenetic tree was constructed using the neighbor-joining method with MEGA 11.0 software, with a bootstrap value of 1000.
Quantitative Real-Time PCR
To validate the structural gene expression profiles, ABScript Neo RT Master Mix for quantitative real-time polymerase chain reaction (qRT-PCR) with gDNA Remover kit was used to reverse transcribe the sample RNA into cDNA.Subsequently, a BrightCycle Universal SYBR Green qPCR Mix with UDG was used for qRT-PCR analysis of DEGs using the Applied Biosystems TM QuantStudio TM 3&5 (Waltham, MA, USA) real-time quantitative PCR instrument.The following conditions were applied: 37 • C for 2 min, 95 • C for 3 min, and then 40 cycles of 95 • C for 5 s and 60 • C for 30 s, followed by a melt cycle of 95 • C for 5 s and 60 • C for 11 min.β-actin was selected as the internal reference gene [20].
Primers for qRT-PCR were designed using Primer Premier 5, ensuring amplicon lengths of 80-150 bp, GC contents of 40-60%, and Tm values of 50-60 • C. The primer sequences can be found in Table S2.All qRT-PCR experiments were taken in three biological replicates, and each reaction was performed in triplicate.The relative expression of each gene was calculated with the 2 −∆∆Ct method [29].
Determination of Puerarin in P. montana var. thomsonii at Different Growth Stages via HPLC
The contents of puerarin in 12 groups of samples were separately determined using HPLC.It was found that the content of puerarin was higher during the vigorous growth stage than during the dormant stage, both in one-year and two-year-old root tubers (Table 2).The raw data of the content determination and the related chromatograms are displayed in Table S1 and Figure S1.To better understand the nutritional and medicinal differences among P. montana var.thomsonii in the two growth stages, widely targeted UPLC-MS/MS-based metabolite profiling of the samples was performed.A total of 1525 metabolites were detected based on the UPLC-MS/MS detection platform and a self-built database (Figure 1a), including 285 flavonoids, 126 lipids, 204 amino acids and derivatives, 152 alkaloids, 76 organic acids, 177 terpenoids, 55 nucleotides and derivatives, 57 lignans and coumarins, 143 lignans and coumarins, 20 quinones, 13 steroids, 4 tannins, and 105 others metabolites [30].The detailed information of these metabolites is shown in Table S3.Of the 285 flavonoids in the four groups, 10 chalcones, 24 flavanones, 3 flavanonols, 1 dihydroisoflavone, 6 anthocyanidins, 77 flavones, 39 flavanols, 96 isoflavones, and 29 other flavonoids were detected.The results of PCA analysis revealed a clear separation between the four groups of samples and quality control (QC) samples, and the repeated samples in each group were gathered, except sample GX2-3, which showed slight variation, indicating the repeatability and reliability of the experiments (Figure 1b).
We used the supervised method, OPLS-DA, and Student's t-test (value of p < 0.05) to find the metabolites responsible for differences among these four groups.In this study, the OPLS-DA model compared metabolite contents of the stages in pairs to evaluate the differences.The Q2 and R2 values of the two comparison groups exceeded 0.5, demonstrating that the models were stable [31] (Table S3).The DAMs (|Log (FC)| > 1, VIP ≥ 1, and p-value < 0.05) between varieties were screened.
In terms of group comparisons, 281 DAMs in the GX1 vs. GD1 group were observed, with 65 upregulated and 216 downregulated metabolites.Similarly, the GX2 vs. GD2 group exhibited 184 DAMs, including 48 upregulated and 136 downregulated metabolites.In addition, based on KEGG enrichment analysis, the DAMs in the two comparison groups were both mainly enriched in "biosynthesis of secondary metabolites" and "isoflavonoid biosynthesis" (Figure 1c).
Transcriptomic Profiling and Differentially Expressed Gene Screening of P. montana var. thomsonii at Different Growth Stages
After testing the RNA quality, a total of 82.98G high-quality bases were generated, with GC bases accounting for 43.86−44.69%and Q30 bases exceeding 92.28%.The RNA integrity was good, and the total amount complied with the requirements for standard library construction.Transcriptome sequencing was performed.PCA separated the two samples in each comparison group clearly (Figure 2a).Differential genes were screened for |log2FC| ≥ 1 and FDR < 0.05.A total of 16300 DEGs were obtained from four groups of samples.There were 7543 DEGs in the GX1 vs. GD1 group, with 3785 upregulated and 3758 downregulated genes.Similarly, the GX2 vs. GD2 group exhibited 7171 DEGs, including 3102 upregulated and 4069 downregulated genes.Detailed information is shown in Table S5.The volcano map visualizes the overall distribution of differential genes in the two groups of samples (Figure 2b), indicating significant differences in gene expression levels during root development in P. montana var.thomsonii.To further analyze the DEGs, KEGG enrichment analysis was conducted, and the top 20 enriched pathways are shown in the bubble diagram.Notably, twelve of these pathways were shared by both comparison groups and primarily enriched in "metabolic pathways" and "starch and sucrose metabolism" (Figure 2c).We used the supervised method, OPLS-DA, and Student's t-test (value of p < 0.05) to find the metabolites responsible for differences among these four groups.In this study, the OPLS-DA model compared metabolite contents of the stages in pairs to evaluate the differences.The Q2 and R2 values of the two comparison groups exceeded 0.5, demonstrating that the models were stable [31] (Table S3).The DAMs (|Log (FC)| > 1, VIP ≥ 1, and p-value < 0.05) between varieties were screened.
thomsonii at Different Growth Stages
After testing the RNA quality, a total of 82.98G high-quality bases were generated, with GC bases accounting for 43.86−44.69%and Q30 bases exceeding 92.28%.The RNA integrity was good, and the total amount complied with the requirements for standard library construction.Transcriptome sequencing was performed.PCA separated the two samples in each comparison group clearly (Figure 2a).In the transcriptome generated in this study, a total of 4479 transcripts were annotated as transcription factors (TFs), mainly belonging to the AP2/ERF-ERF, WRKY, MYB, bHLH, C2H2, SNF2, and C3H families (Figure S2).MYB transcription factors have been shown to regulate secondary metabolism, stress responses, and development in various plants [32].Specifically, in P. montana var.thomsonii, they play a role in the biosynthesis of puerarin, where the expression of nine key enzymes is regulated by specific MYBs in the phenylpropanoid and isoflavonoid pathways [33].A total of 261 MYBs were detected in this study.
Analysis of Metabolites and Differentially Expressed Genes Involved in the Puerarin Biosynthesis Pathway
Based on previous research by Shang et al. [21], a possible pathway map for puerarin biosynthesis was constructed.The conversion of phenylalanine to cinnamic acid is catalyzed by PAL, followed by the conversion of cinnamic acid to 4-coumarin coenzyme A through the actions of C4H and 4CL.Subsequently, CHS and CHR polymerize 4-coumarin coenzyme A to isoliquiritigenin.CHI then catalyzes the formation of liquiritigenin from isoliquiritigenin, which is further catalyzed by IFS to produce 2,7,4 ′ -trihydroxyisoflavonone.Thereafter, chalcone isoflavone can be catalyzed by HID and 8-C-GT to produce the target compound, puerarin, via two different pathways (Figure 3).Among all the metabolites, six metabolites were labeled in the puerarin biosynthesis pathway, which were puerarin, daidzin, daidzein, isoliquiritigenin, liquiritigenin, and Lphenylalanine [34].The expression levels of these metabolites were analyzed using a heat map, as shown in Figure 3.As expected, in a two-by-two comparison, the downstream metabolites (puerarin and daidzin) accumulated more in the vigorous growth group, whereas the upstream compounds (isoliquiritigenin, liquiritigenin, and L-phenylalanine) accumulated more in the dormant group.In particular, differences in upstream metabolites were more pronounced in the one-year samples, while differences in downstream metabolites were more pronounced in the two-year samples.
Using profile hidden Markov models (HMMERs) and local BLAST 2.14.0 software, several key structural genes involved in the puerarin biosynthesis pathway were identified, including 12 PALs, 109 4CLs, 53 CHSs, 92 CHRs, 11 CHIs, 19 IFSs, 510 UGTs, and 91 HIDs.All DEGs were combined with all structural genes on the pathway to form an intersection, and the genes in common are likely to be differential structural genes involved in the puerarin biosynthesis pathway.In the GX1 vs. GD1 group, a total of 114 DEGs can Among all the metabolites, six metabolites were labeled in the puerarin biosynthesis pathway, which were puerarin, daidzin, daidzein, isoliquiritigenin, liquiritigenin, and L-phenylalanine [34].The expression levels of these metabolites were analyzed using a heat map, as shown in Figure 3.As expected, in a two-by-two comparison, the downstream metabolites (puerarin and daidzin) accumulated more in the vigorous growth group, whereas the upstream compounds (isoliquiritigenin, liquiritigenin, and L-phenylalanine) accumulated more in the dormant group.In particular, differences in upstream metabolites were more pronounced in the one-year samples, while differences in downstream metabolites were more pronounced in the two-year samples.
All DEGs were combined with all structural genes on the pathway to form an intersection, and the genes in common are likely to be differential structural genes involved in the puerarin biosynthesis pathway.In the GX1 vs. GD1 group, a total of 114 DEGs can be labeled on the pathway, including 3 PALs, 6 IFSs, 8 4CLs, 4 CHSs, 16 CHRs, 1 CHI, 70 UGTs, and 6 HIDs.Similarly, in the GX2 vs. GD2 group, a total of 118 DEGs can be labeled on the pathway, including 7 IFSs, 10 4CLs, 3 CHSs, 12 CHRs, 2 CHIs, 76 UGTs, and 8 HIDs.Notably, there were 67 common DEGs between the two groups, including 7 4CLs, 2 CHSs, 10 CHRs, 3 IFSs, 43 UGTs, and 2 HIDs.A heatmap of gene expression was derived based on FPKM values, as shown in Table S5, and it is also labeled next to each step in Figure 3.As can be seen from the data, the expression of UGTs has a certain regularity.That is, the expression was higher at the vigorous growth stage (GX) than at the dormant stage (GD), and the expression was especially high at the vigorous growth stage during the second year (GX2).IFS was also expressed significantly higher at the vigorous growth stage than at the dormant stage.Similarly, HID was also expressed at high levels at the vigorous growth stage, but the differences were not as pronounced.
Phylogenetic Analysis of Candidate UGTs
To further evaluate their possible regulatory roles, a phylogenetic analysis was conducted on the 43 UGTs, along with the previously characterized UGTs of known structures and functions from other species [35,36].Orthologous genes with similar functions are usually clustered in the same clades (Figure 4).The information on UGTs used in phylogenetic analysis is listed in Table S6.Based on the phylogenetic tree topology, the proteins were clustered into nine clades.Only PtUGT35 was not assigned to any clades.Except for 14 UGTs clustered together that did not cluster with previously reported UGTs, the remaining 28 UGTs all clustered with different UGTs, which can be differentiated into 17 OGTs and 11 CGTs.A total of 17 OGTs can be further subdivided into 8 isoflavone 7-O glycosylations, 5 flavone 5-O glycosylations, and 4 flavone 7-O glycosylations.Notably, the eight isoflavone 7-O glycosylations (PtUGT1, PtUGT2, PtUGT3, PtUGT4, PtUGT7, PtUGT8, PtUGT32, and PtUGT36) clustered with 7-O-GT reported in P. montana var.lobata might be involved in catalyzing the production of daidzein from daidzein [37].Four of the eleven CGTs (PtUGT33, PtUGT34, PtUGT38, and PtUGT39) clustered with PlUGT43, which is the first 8-C-GT proven to catalyze the generation of puerarin from daidzein in puerarin biosynthesis [22], and thus, it can be hypothesized that these four 8-C-GTs in the present study also have the same function, pending further verification.The other seven clustered with the 8-C-GT of Trollius chinensis [38].The protein sequences of candidate UGTs are listed in Table S7.
Co-Expression Analysis of Important Genes and Metabolites
In order to examine the relationship between important metabolites and genes related to puerarin biosynthesis in P. montana var.thomsonii at different growth stages, a coexpression network analysis of chosen metabolites and genes was conducted.Correlations were analyzed, and network diagrams were drawn based on the Pearson correlation coefficient, as well as the p-value (Figure 5).The color of the line indicates the correlation, where red is a positive correlation and blue is a negative correlation; the thickness of the line indicates the size of the p-value; and the metabolites and genes are differentiated by the shapes of the nodes.It can be observed that as an important intermediate, daidzein is strongly associated with almost all candidate UGTs and shared differential HID genes.Moreover, puerarin is tightly correlated with the four 8-C-GTs and less correlated with the two HIDs, which also corroborates the pathway by which daidzein is catalyzed by 8-C-GT to produce puerarin.The associations between daidzein and puerarin for these candidate UGTs are roughly consistent.
of the eleven CGTs (PtUGT33, PtUGT34, PtUGT38, and PtUGT39) clustered with PlUGT43, which is the first 8-C-GT proven to catalyze the generation of puerarin from daidzein in puerarin biosynthesis [22], and thus, it can be hypothesized that these four 8-C-GTs in the present study also have the same function, pending further verification.The other seven clustered with the 8-C-GT of Trollius chinensis [38].The protein sequences of candidate UGTs are listed in Table S7.
Co-Expression Analysis of Important Genes and Metabolites
In order to examine the relationship between important metabolites and genes related to puerarin biosynthesis in P. montana var.thomsonii at different growth stages, a co-expression network analysis of chosen metabolites and genes was conducted.Correlations were analyzed, and network diagrams were drawn based on the Pearson correlation coefficient, as well as the p-value (Figure 5).The color of the line indicates the correlation, where red is a positive correlation and blue is a negative correlation; the thickness of the line indicates the size of the p-value; and the metabolites and genes are differentiated by the shapes of the nodes.It can be observed that as an important intermediate, daidzein is strongly associated with almost all candidate UGTs and shared differential HID genes.Moreover, puerarin is tightly correlated with the four 8-C-GTs and less correlated with the two HIDs, which also corroborates the pathway by which daidzein is catalyzed by 8-C-GT to produce puerarin.The associations between daidzein and puerarin for these candidate UGTs are roughly consistent.
Quantitative Real-Time Polymerase Chain Reaction Validation of the Expression Levels of the Genes Associated with Puerarin Biosynthesis
To confirm the RNA-Seq results, the transcript abundance of eight genes in each of the two comparison groups was analyzed via qRT-PCR [39].These genes included the main structural genes of the flavonoid biosynthetic pathway.The results showed that the expressions of the genes related to puerarin biosynthesis determined via qRT-PCR were
Quantitative Real-Time Polymerase Chain Reaction Validation of the Expression Levels of the Genes Associated with Puerarin Biosynthesis
To confirm the RNA-Seq results, the transcript abundance of eight genes in each of the two comparison groups was analyzed via qRT-PCR [39].These genes included the main structural genes of the flavonoid biosynthetic pathway.The results showed that the expressions of the genes related to puerarin biosynthesis determined via qRT-PCR were consistent with the corresponding FPKM values obtained from the RNA-Seq analysis (Figure 6).
Discussion
In recent years, the advancement of metabolomics and transcriptomics has provided new approaches for biosynthetic mechanisms of secondary metabolites in medicinal plants.Therefore, further studies have been conducted on the puerarin biosynthesis pathway in Pueraria [1,9,14,19,40].While the upstream of the pathway has been well characterized, the enzymes responsible for synthesizing the key downstream metabolite puerarin are still controversial and subject to some contradictions.Moreover, most of these studies focus on comparing transcriptomes of different tissue parts, such as roots,
Discussion
In recent years, the advancement of metabolomics and transcriptomics has provided new approaches for biosynthetic mechanisms of secondary metabolites in medicinal plants.
Therefore, further studies have been conducted on the puerarin biosynthesis pathway in Pueraria [1,9,14,19,40].While the upstream of the pathway has been well characterized, the enzymes responsible for synthesizing the key downstream metabolite puerarin are still controversial and subject to some contradictions.Moreover, most of these studies focus on comparing transcriptomes of different tissue parts, such as roots, stems, and leaves, or comparing different varieties to identify candidate genes that play roles in the puerarin biosynthesis pathway.However, few studies of P. montana var.thomsonii have been conducted to investigate the comparative transcriptomic and metabolomic studies to probe the differential expression of structural genes and accumulation of metabolites on puerarin biosynthesis pathways from the perspective of growth stages.
The upstream phenylalanine pathway of the puerarin biosynthesis has been well studied, but the downstream isoflavonoid pathway remains controversial, especially the stage at which the 8-C glycosylation reaction occurs.Previous research has suggested two possible pathways.Puerarin can occur via a route where the C-glycosyl linkage is introduced to the chalcone isoliquiritigenin or through the direct action of 8-C-GT on daidzein.However, no differential expression data of the puerarin precursor 2,7,4 ′ -trihydroxyisoflavanone 8-C-glycosyl, which supported the first pathway, were obtained in our study.However, we detected differential accumulation data of daidzein, which further validated the introduction of 8-C glycosyl into the daidzein pathway.Additionally, in our study, we found some common patterns in the accumulation of the metabolites on the puerarin pathways.Differences in the accumulation of upstream and downstream metabolites in the pathway at different growth stages then lead to the regularity of accumulation stages for isoflavone biosynthesis.The accumulation of upstream metabolites was mainly accomplished during the dormant stage, while the accumulation of downstream target products occurred mainly during the vigorous growth stage.Moreover, it is worth noting that the difference in the accumulation of upstream compound metabolites was more obvious in the comparison of the GX1 vs. GD1 group, and the difference in the accumulation of downstream target compounds was more obvious in the comparison of the GX2 vs. GD2 group, so it can be speculated that the upstream metabolites that accumulated in the first year were consumed in the second year to obtain the downstream target products.Therefore, this study also provides an information basis for further exploring and verifying the impact of age on the puerarin biosynthesis pathway.
Glycosyltransferase (GT) is a superfamily of proteins that catalyze glycosylation modification reactions, which can alter the solubility, stability, and other properties of the substrate to make the substrate molecule more versatile [41].Among all the GT families, the UGT family is most closely related to the downstream metabolites puerarin and daidzin.In our study, the 43 PtUGTs from both comparison groups were used for phylogenetic analysis.The results revealed that eight 7-O-GTs might be involved in the catalytic synthesis of daidzin from daidzein; in addition, four 8-C-GTs (PtUGT33, PtUGT 34, PtUGT 38, and PtUGT 39) clustered closely with the previously reported PlUGT43, which demonstrates that they might play an important role in catalyzing the generation of puerarin from daidzein [22,37].Therefore, these four UGTs could be the most likely candidates.It is further evidence of the authenticity of the pathway that the C-glycosyl linkage is introduced to daidzein to produce puerarin.However, their function cannot be determined based on sequence alone, and whether they play a role in puerarin biosynthesis remains to be further verified.
All these studies contribute to the knowledge of puerarin biosynthesis and offer information about the stage at which the 8-C-glycosylation reaction occurs in the pathway.These findings provide valuable insights into the cultivation and quality enhancement of Pueraria montana var.thomsonii.
Conclusions
In our study, P. montana var.thomsonii from two growing years, at both the vigorous and dormant growth stages, were selected to explore the underlying mechanism of the puerarin biosynthesis pathway, from both phenotypic and genetic perspectives.The results of the metabolic analysis showed that a total of 1525 metabolites were detected and quantified, of which flavonoids were dominant.Transcriptomic analysis showed that a total of 16,300 DEGs were identified, of which 67 DEGs were common between the two groups in the puerarin biosynthesis pathway.The accumulation stage of the major metabolites in the pathway was investigated, and we found that the accumulation of upstream metabolites is mainly accomplished during the dormant stage, whereas the accumulation of downstream target products occurs mainly during the vigorous growth stage.This research provides new guidance for the cultivation of P. montana var.thomsonii.The genes involved in the isoflavone biosynthesis pathway were further refined in our studies, identifying four 8-C-GTs that might be able to act on daidzein to generate puerarin and eight 7-O-GTs that might be able to act on daidzein to generate daidzin.These findings further validated the authenticity of the daidzein pathway to obtain puerarin.The integrated transcriptomic and metabolomic analyses provide evidence for the C-glycosylation stage of the puerarin biosynthetic pathway and the mechanism of metabolite accumulation and also provide a reference for the quality enhancement of P. montana var.thomsonii.
Genes 2023 , 17 Figure 1 .
Figure 1.Metabolite accumulation of P. montana var.thomsonii at different growth stages.(a) Metabolite accumulation at different growth stages.(b) Principal component analysis (PCA) of metabolites in different groups.(c) KEGG analysis of all DAMs: the upper section represents the results of the GX1 vs. GD1 group, while the lower section represents the results of the GX2 vs. GD2 group.Each bubble in the plot represents a metabolic pathway, the abscissa and bubble size of which jointly indicate the magnitude of the impact factors of the pathway.A larger bubble size indicates a larger impact factor.The abscissa of bubbles indicates the enrichment ratio of metabolites in each pathway.The bubble colors represent the p-value of the enrichment analysis, with lighter colors showing a higher confidence level.Sorted by the p-value, the top 20 metabolic pathways are plotted in the bubble chart.
Figure 1 .
Figure 1.Metabolite accumulation of P. montana var.thomsonii at different growth stages.(a) Metabolite accumulation at different growth stages.(b) Principal component analysis (PCA) of metabolites in different groups.(c) KEGG analysis of all DAMs: the upper section represents the results of the GX1 vs. GD1 group, while the lower section represents the results of the GX2 vs. GD2 group.Each bubble in the plot represents a metabolic pathway, the abscissa and bubble size of which jointly indicate the magnitude of the impact factors of the pathway.A larger bubble size indicates a larger impact factor.The abscissa of bubbles indicates the enrichment ratio of metabolites in each pathway.The bubble colors represent the p-value of the enrichment analysis, with lighter colors showing a higher confidence level.Sorted by the p-value, the top 20 metabolic pathways are plotted in the bubble chart.
Figure 2 .
Figure 2. Analysis of differentially expressed genes (DEGs) of P.montana var.thomsonii at different growth stages: the left side of the figure represents the GX1 vs. GD1 group, while the right side
Figure 2 .
Figure 2. Analysis of differentially expressed genes (DEGs) of P.montana var.thomsonii at different growth stages: the left side of the figure represents the GX1 vs. GD1 group, while the right side represents the GX2 vs. GD2 group.(a) Principal component analysis (PCA) of the two comparison groups.(b) Differential gene volcano maps for the two comparison groups.(c) KEGG analysis of all DEGs for the two comparison groups.Each bubble in the plot represents a pathway, the abscissa and bubble size of which jointly indicate the magnitude of the impact factors of the pathway.A larger bubble size indicates a larger impact factor.The abscissa of bubbles indicates the enrichment ratio of genes in each pathway.The bubble colors represent the Q-value of the enrichment analysis, with lighter colors showing a higher confidence level.Sorted by the Q-value, the top 20 metabolic pathways are plotted in the bubble chart.
Genes 2023 , 17 Figure 3 .
Figure 3. Puerarin biosynthetic pathways in P. montana var.thomsonii.DEG heatmaps are labeled at each step, and the DAM heatmap is labeled in the lower right corner.
Figure 3 .
Figure 3. Puerarin biosynthetic pathways in P. montana var.thomsonii.DEG heatmaps are labeled at each step, and the DAM heatmap is labeled in the lower right corner.
Figure 4 .
Figure 4. Phylogenetic analysis of 43 differential UGT genes shared by the two comparison groups.The red-labeled UGTs are highly homologous to 8-C-GT UGT43 in P. montana var.lobata, and yellow-labeled UGTs are highly homologous to 7-O-GT PlUGT1 in P. montana var.lobata.
Figure 4 .
Figure 4. Phylogenetic analysis of 43 differential UGT genes shared by the two comparison groups.The red-labeled UGTs are highly homologous to 8-C-GT UGT43 in P. montana var.lobata, and yellow-labeled UGTs are highly homologous to 7-O-GT PlUGT1 in P. montana var.lobata.
Figure 5 .
Figure 5. Interaction network of important differential structural genes and metabolites involved in puerarin biosynthesis.Circles indicate metabolites, diamonds indicate structural genes, and different genes are distinguished by different colors, where 8-C-GT is dark blue, 7-O-GT is light blue, and HID is gray.
Figure 5 .
Figure 5. Interaction network of important differential structural genes and metabolites involved in puerarin biosynthesis.Circles indicate metabolites, diamonds indicate structural genes, and different genes are distinguished by different colors, where 8-C-GT is dark blue, 7-O-GT is light blue, and HID is gray.
Figure 6 .
Figure 6.qRT-PCR validation of the genes associated with puerarin biosynthesis in P. montana var.thomsonii at different growth stages.The left is the one-year-old tuber comparison group, and the right is the two-year-old tuber comparison group.Values are means ± SD of three biological replicates (n = 3).** p < 0.01, * p < 0.05, n.s.-not significant.
Figure 6 .
Figure 6.qRT-PCR validation of the genes associated with puerarin biosynthesis in P. montana var.thomsonii at different growth stages.The left is the one-year-old tuber comparison group, and the right is the two-year-old tuber comparison group.Values are means ± SD of three biological replicates (n = 3).** p < 0.01, * p < 0.05, n.s.-not significant.
Table 1 .
Sample information of P. montana var.thomsonii.
Table 2 .
Puerarin content connected with sample information. | 8,829.8 | 2023-12-01T00:00:00.000 | [
"Biology",
"Environmental Science"
] |
AVIGILON COMPACT CAMERA’S TEST FOR INTEGRATED SAFETY SYSTEM WITHIN AIRPORT SECURITY
The article presents an experimental exploration of the selected technical features of the Avigilon 2.0C-H4A-BO1-IR Compact Zoom Camera with IR Adaptive Illumination. The article describes the purpose, procedure, and results of the motion detection verification, as well as the identification of motion detection errors, using Avigilon's investigated camera, to the distance of guaranteed recognition capability in specific daylight conditions that determines video analysis. This article constitutes the first part of the internal research activity of the Department of Flight Preparation pre-research, for the design of an integrated mobile airport security system. For safety reasons, the testing was performed near the airport and not at the airport. The test sample was obtained by using the Avigilon 2.0C-H4A-BO1-IR camera located 8 meters above the ground level in the direction of the selected perimeter of the "protected area" for the experiment. The movement in the space was made by people and the passage of motor vehicles at a distance that was less than the distance guaranteed by the camera's recognition capability in the specific daylight conditions. The movement of persons and motor vehicles was generally performed perpendicular to the position of the camera, left to right, and/or back. The speed of movement of people was, as a rule, an average walking speed of 1m/s, the motor vehicles ranging up to 40km /h. Identification of motion detection errors is important for corrections of the prepared information model of security risk assessment of a protected object based on the fuzzy logic to support the airport security management decisions, as well as for finding a technical solution to eliminate these camera vulnerabilities, or selecting and testing another camera for our mobile technology platform. The results advance our theoretical knowledge and have a praxeological significance for the creation of a technological demonstrator and subsequently a prototype of a smart mobile airport security system. Institutions responsible for the protection of state borders, the fight against illegal migration, smuggling of goods, etc. are also interested in mobile security solutions.
Introduction
Based on the definition of state security in the Constitutional Act of the Slovak Republic no. 227/2002 Collection of Laws we can identify the following protected interests: ensuring the sovereignty, territorial integrity and the inviolability of state borders; protection of the lives and human health; property protection; the protection of fundamental human rights and freedoms, the protection of economic interests and the environment. An important part of the protected interests is also the critical infrastructure in which the "Transport" sector plays a key role. To protect interests, we create social and complex technical systems for working in the digital and realworld space. In the technogenic security sector, we are exploring − safety and the human factor in technical and technological processes, − Critical Infrastructure Protection / Critical Infrastructure sectors and subsectors, − information protection, cyberspace protection, cybercrime, − security challenges of science and technology development for an individual level of security − citizen safety, protected rights and interests, and protection of the population, etc. (Kelemen, 2017). Real-time protection of interests and predictions are possible if they are linked to a real-time data provider using modern information and communication technologies, including the camera systems, image analysis tools, sensors, etc. We find inspirational applications at work as on the analysis of spatiotemporal data to predict traffic conditions (Kyriakou et al., 2019), an integrated approach to information analysis (Semenov et al., 2019) or on the analysis of the traffic stream distribution, etc. (Ambroziak et al., 2014). The description of the current state of knowledge presented in world literature in relation to the research area presented in article on the selected tools for the object or critical infrastructure protection and critical infrastructure failure and risks we can also study generally in the agenda of the European Data protection Supervisor before the user acceptance test, in the image quality testing (IEC, 2015), or in the work on the specification of operating systems of Internet of Things cameras (Palmer, 2017), on safety and security issues in technical infrastructures (Rehak et al., 2020), on the technology agenda within the critical infrastructure and integrated protection (Vidrikova et al., 2017). We can be inspired by the work of CCTV System Performance Specifications (Young, 2015), on discussion of the requirements and applications of the Airport Security Program (Price, Forrest, 2016), or discussion on perimeter security and lists and defines several types of barriers (Purpura, 2017). After system studies, we find an examination of the specific properties of sensors on owing to a lack of computational resources, lightweight and computationally efficient face recognition algorithms are required (Lee et al., 2020), also the many fusion methods for information acquired from sensors (cameras) that have been proposed in the literature for activity recognition (Aguileta et al., 2019), etc. For the Transport sector/subsector Air transport is a lasting issue of ensuring the safety of air traffic, the protection of persons and property, which affects the air transport and shipping processes. Creating the integrated security systems for airports is one of the available and effective tools to handle the certain risks of the required state. Engineers and researchers are challenged to research, and to develop an innovate relevant technology, collect, process, and visualize data to ensure the security of protected interest as in another works (Szabo et al., 2019) and (Tobisová et al., 2017). The camera systems are such effective tools whose technical parameters and intelligent capabilities are subject to investigation and verification. The purpose of the pre-research is to support the decision-making process of the end user on the selection of sensors (cameras) for the mobile airport security project for helicopters, based on experimental verification of digital cameras from pre-selection, in accordance with input specifications and customer requirements. The Avigilon and Hikvision cameras meet the entry requirements and therefore 2 independent verifications were performed in the standard situation before the verification of the cameras in non-standard situations and before the subsequent verification within the so-called user acceptance test. The article aims to present the selected outputs of the motion detection verification experiment, as well as the identification of motion detection errors using the Avigilon 2.0C-H4A-BO1-IR camera to the distance of the guaranteed recognition ability under the specific daylight conditions that determines the video analysis.
The research methodology of the article was based primarily on quantitative methodology. The nature of the tasks required the implementation of pre-experimental research on the issue, such as case study research design, which takes into account one dependent variable. The evaluation criterion was the number of errors in the detection of objects within the monitored space, in the specified time and monitoring conditions. We assume the thesis that the results of the verified Avigilon camera will not show errors in the detection of persons or in the incomplete display of a person, but will show errors in the detection of vehicles as objects with a higher speed of movement. For these reasons, the two basic research questions were formulated for the planned pre-experimental research of the issue: -what is the number of errors in detecting movement of people and vehicles from the test image sample in a defined "protected area" for the electronic monitoring using the Avigilon 2.0C-H4A-BO1-IR camera that we can identify? -to find out whether an analytical tool (camera, SW) can only register a human being when the whole body is clearly visible or evaluating an object as a human being even when the person's personality is not visible?
Research data and description of functional
verification Experimental research is a scientific approach to research, where one or more independent variables are manipulated and applied to one or more dependent variables to measure their effect on the latter (https://www.formpl.us/blog/experimental-research). Causality is the core of the experiment. The researchers observed, described, measured, scaled the results and evaluated the changes according to the digital recording and detection of the phenomenon, the causality of which they monitored during our pre-experimental research. We have two samples of digital IP cameras (Avigilon, Hikvision), which we expose to simultaneous recording of the same situations, at the same time and in the field of surveillance. They are treated the same. After research, we will find out which camera shows more cases of motion detection error identification. Therefore, we can come to the partial conclusion that which camera shows better parameters in standard situations, in accordance with the customer's input requirements. The standard situation is the movement of people, vehicles and other subjects in the observation area, where the researcher influences the time and space of the camera's surveillance, but not the number of subjects for editing and monitors the consequences. In the following experiment, in non-standard situations, we calculate, examine and compare the achieved results (identify, analyze and evaluate). The non-standard situation will be the movement of persons, vehicles and other entities in the monitored area, where the researcher influences the variables according to different scenarios of the tactical situation and conditions, and monitors their consequences. It will seek to make it more difficult to work with cameras (masked persons, deliberate maneuvers of persons in the field of surveillance, daylight, night light, drones, luggage, animals in space, etc.). The article presents the results of a case study of the Avigilon 2.0C-H4A-BO1-IR Compact Zoom camera. The second article will present the results of a case study of a Hikvision camera. Examining the selected technical parameters of the Avigilon 2.0C-H4A-BO1-IR Compact Zoom Camera was a work with a large number of digital data, which was the main material for the experiment. The database has been searched for and processed based on certain rules. Before designing our experiment, we studied the similar procedures, findings and work of researchers who were somewhat concerned with the camera systems. The expert community confirms the experience that the video motion detection required by us is the basic and prevailing video analysis method in camera systems (Caputo, 2014; Loveček and Reitšpís, 2011; STN, 2013; 2013; STN, 2014). For example, you can track the line crossing, person counting, image motion, object tracking, neural networks, pixel motion detection, or the classic motion detection, etc. To achieve the goal, camera-based camera assessment (object identification, analysis, and evaluation) was compared with the detection of the analytical SW tool within the following Avigilon 2.0C-H4A-BO1-IR camera experimentation (the experimental limitations): − selecting a "protected area" perimeter for the experiment, − realization of the design of the location of the camera, − realizing the process of "camera learning" in a selected perimeter of the protected area, − the trial operation of monitoring and recording of motion in the selected "protected area" area for the experiment, approximately within 2 hours, − "experimental operation" of monitoring and recording of movement in the selected experimental area of the "protected area" for the experiment, approximately 5 hours, − checking collection, storing records in the archive, within 5 hours, − identifying and selecting the "experimental sample record" in the selected timeframe from the scanned "protected area", − performing digital image analysis based on motion detection of persons and vehicles in the selected "experimental sample record", − verifying the level of success in detecting the movement of persons and motor vehicles, − identification and number of errors in the detection of movement of persons and motor vehicles, in the selected "experimental sample record", − verification of digital data from the selected "experimental sample record", − processing the discussion of the results of the experiment for the Avigilon 2.0C-H4A-BO1-IR camera, as part of the pre-survey. An experimental sample of the record was selected from the database of digital data obtained during the 3 days of the pre-experiment. The test sample is from the 15th of October 2018, at 12:30-13:30, using the Avigilon 2.0C-H4A-BO1-IR camera located 8 meters above the ground level in the direction of the selected perimeter of the "protected area" for the experiment. The movement in the space was made by people and the passage of motor vehicles at a distance that was less than the distance guaranteed by the camera's recognition capability in the specific daylight conditions. The movement of persons and motor vehicles was generally performed perpendicular to the position of the camera, left to right, and/or back. The speed of movement of people was, as a rule, an average walking speed of 1m/s, the motor vehicles ranging up to 40km /h. During the pre-experiment, the privacy policy was followed.
Results and discussion on Avigilon experiment 1_2018
The pre-experiment database was created from the 3 days from 14th to 16th October 2018. A comparison of the image analysis was performed on the selected test image, dated 15.10.2018, at 12: 30-13: 30. When comparing the video-analysis, only systems using the so-called "qualified video-analysis, that is, that the system knows with a certain degree of probability that a person is present in the picture, respectively the car. No pixel-based video-analytic functions were also used, which only track the changes of individual pixels in a given image without being able to determine qualitatively what to engage in the records. We do not mean that a simpler, the pixel analysis is not usable, but for our experiment, we have focused on a qualified form of analysis. We left all the systems in the automatic setting, so the values for the sensitivity, the size of the analyzed objects, the speed of movement, and so on, were left to set the system on its own. We have learned from our previous work on resolving crises and protecting critical infrastructure also (Kelemen et al., 2014(Kelemen et al., , 2015.
Camera system Avigilon 2.0C-H4A-BO1-IR By camera from Avigilon we found that the camera qualified the 63 people out of a total of 75 in the test records. We have manually searched for people who have not marked as a qualified object of interest and have verified why they have not registered there by the system. The result of the analysis is the finding that the 11 people were the case when two or more people crossed each other simultaneously. The probability of object detection has reached 84%.
Motion Detection Analysis -1
The test records were prepared to demonstrate the capabilities and the current state of the video-analytics tools of the camera. At the top of the picture there is a sidewalk that is interrupted by the standing cars, with the pedestrians passing through the scene alternately showing the whole or only the upper part of the person's body. The purpose of the experiment was to find out if an analytical tool (camera, SW tool) can be registered with a human when the whole body is visible or evaluating the object as a human being even when the body is not visible. At the bottom of the scene there is a roadside parking where it is possible to track the cars passing by the car at different speeds and directions. Due to the fact that it is a parking place, it is the only a speed in the range of 0 to 40 km/h. In the test camera, we deliberately used the longer focal length of the lens, that is, the narrower angle of the shot, so that the passing objects, whether the cars or people, were shot at relatively short distances, thereby placing the higher demands on the success and accuracy of the video analytics tool. In the first test record, one person who was not registered with the camera, it was the case of a cyclist. In this case, the system did not evaluate the object as a person. The reason may be the fact that the cyclist was not visible as a whole, while at the same time it was at another car (but it was qualified). The speed of movement seemed subjective to a qualified label. The car of dark colors is an incomplete image but the camera system detected it (Figure 1.).
Motion Detection Analysis -2
When detecting the cars as a qualified object, it is rather difficult to determine why the system did not know identified the car of the white color as an object. One of the possible causes may be a fast-moving car that has not been in the field of vision for a long enough time to identify it as a vehicle. The second reason can be the fact that cars often went so that they were not visible as a whole in the camera's field of view (Figure 2.).
Motion Detection Analysis -3
The experiment demonstrated the fact that the car, as in the previous situation, it was not entirely in the test record, but the camera system qualified it as a car. We believe that in this case, it is a possible reason that the car was of black color, which could help the camera system to the contrast of object to the recognition (Figure 3.).
Motion Detection Analysis -4
Experimental records from the Avigilon camera were also tested and compared with the Axxon Next software tool's analysis. Data from the pre-experimental camera research in standard situations were evaluated within the camera software for analytics. Due to the effort of objective evaluation obtained from the monitored area using a selected camera and identification of errors in object detection, we also used an independent situation analysis detection tool Axxon Next from AxxonSoft (AXXON, 2020).
The success rate of the Axxon Next software was lower than the previous Avigilon system. We think the reason may be that we left the system in the default mode, so we left all the values of the analysis at their original values. At these values, Axxon Next was perceptually relatively successful in the identifying of people, with a qualified 80% of the subjects being shot. However, the basic settings were insufficiently set for a qualified car identification, where the software qualified only 32.56% of cars in the records. It follows from the above that, at the basic setting, the Avigilon system produces better results without the need for additional adjustment of different analysis values, such as external SW. On the other hand, the second comparison system from Axxon Next has to analyze the records before using to create enough time for the "learning" process and set the SW for a specific purpose (Figure 4.) The results of the pre-experiment, followed by the presentation of selected situations in the test records, are important for the research team's knowledge and the creation of the expert database for the following functional verification of Hikvision camera, for their comparison and formulation of the new project under preparation. The Avigilon 2.0C-H4A-BO1-IR pre-imaging camera could be the part of collecting and researching data of the other selected components for the comprehensive ISS C-4 integrated security system design of mobile helicopter airport. C-4 allows the Transport sector to centralize the security performed by both the public and private organizations at the airports, railways, road transport, and ports (SC4, 2020). Discussion on the overall key findings of the Avigilon 2.0C-H4A-BO1-IR camera verification, that relate to established research questions, they have shown that: − the realization of the process of "learning" the camera in the selected perimeter of the "protected area" for the experiment lasted 2.5 hours, − the camera required the further manual refinement of the process of "the camera learning" in a selected "protected area" perimeter, by an experienced operator, − the camera required an additional manual clarification of "the camera learning" from detected own errors in the object detection, − the design and implementation of Avigilon camera placement should not be perpendicular to the planned movement of objects (persons and cars) in the protected area due to the increased time for the object detection process, − in the case of 2 or more persons moving in a space concurrently in a group, they cannot identify the number of persons but they made their detection as 1 object, − in the case of a person with a baby carriage and a child up to 1 m high, the system detected a child carriage as a car identified the person separately but did not identify the child, − if the person was behind the car and a person's shadow was visible on the road, the system detected the shadow of the person on the road. This fact is important in the case of the likely detection of a person hidden behind an object, − the probability of detecting objects using the Avigilon camera reached 84%.
Conclusions
Praxeological experience confirms the general recognition that, when evaluating the system, probability equal to 1 (100% detection) could be considered if the camera system is set and working under an ideal detection condition (Nilsson, 2009). Reallife does not provide the optimal conditions, so it is also necessary to experimentally verify this phenomenon. CCTV manufacturers are looking to build integrated security systems that include their products. An important part of the protected interests is also the critical infrastructure in which "the Transport" sector and "the Air transport subsector" play a key role. For the selection of suitable camera systems and other components of the integrated security system of the airport, it is necessary to obtain the operational data on the capabilities of the cameras in the practical operation from the manufacturer, by the reference from the current system users or by our testing. When we polarize our pre-research thesis and examine its validity, whether we consider causality correctly with respect to the results of the pre-experiment of the Avigilon camera: the results of the verified Avigilon camera will show errors in the detection of persons or in the case of incomplete display of a person, but will not show errors in the detection of vehicles as objects with a higher speed of movement, then we state: the results of the verified Avigilon camera show errors in the detection of persons, especially in the case of an incomplete image of a person, but do not show errors in the detection of cars as objects with a higher speed.
According the results we register errors in the detection of people in dark clothes, if they were on a bicycle. We do not detect errors in the detection of persons in light clothing if they were on a bicycle. Detection of people as pedestrians in various clothes was without problems. Contrary to this finding, we register a white (or light blue) car detection error when moving perpendicular to the camera scanning direction, but we do not register a black car detection error when moving perpendicular to the camera scanning direction. We will use this causality and knowledge in the preparation of the continuation of the functional verification of the Avigilon and Hikvision cameras in non-standard tactical situations using the method of creating scenarios with other variables. The ultimate goal is to select and recommend the best sensor (camera) to perform the task in the mobile airport security system. Our initial research thesis was not confirmed, but proved the interrelationship (causality) for errors in the detection of objects with different color gamut, in the detection of people and other objects, which will significantly affect the creation of scenarios for further functional verification of the camera Avigilon, resp. also Hikvision cameras, in more complex variable conditions for detection. The results of the functional verification of the Avigilon camera in pre-experimental conditions reached a level of 84% certainty and success for object detection. The pre-experiment data have enabled us to respond to the identified research questions and it will be used in the comparison to the other integrated security system components. The pre-experiment results met our expectations, when the camera was originally set up. To clarify further research on the object detection and to find the optimum camera placement, we plan another Avigilon / Hikvision experiments 2020. Future research work will aim to collect the additional data into the C-4 expert database, with a tactical subject experiment. The objects will intentionally move and steal in a protected area before detecting a skewed camera. Data will be used to perform a comparison with other selected cameras, and so on. Part of C-4's integrated security system of the mobile airport will include the recommendations for the system operators, based on our test results and the error detection in the object detections. The importance of the first pre-experiment results of the Avigilon camera stems from the following preexperiment of the Hikvison camera, comparison of their results and recommendation for the end user to the prepared project of a complex design of a mobile airport for helicopter operation with the following partial solutions: − a lockable mobile hangar, − the mobile shelter for the helicopter with the technical containers for staff and material, − the mobile backup power source, − the innovative mobile transport platform for the small and medium-sized helicopters, − the portable light technology system for the airport, a heliport, an operating part of the airport, − the physical and object protection of airport and integrated mobile airport security system, − the research of the complex communication layer security, secure and controlled WIFI environment, which is the part of the mobile airport infrastructure, − the design of mobile infrastructure for the airline staff, passengers, cargo, etc., − the visualization of the outputs and design of the mobile airport, − the project documentation for the technical design of devices/products will also be output. This article constitutes the first part of the research activity of the Department of Flight Preparation as the pre-research, for the design of an integrated mobile airport security system. For safety reasons, the testing was performed near the airport and not at the airport. Identification of motion detection errors is important for corrections of the prepared information model of security risk assessment of a protected object based on the fuzzy logic to support the airport security management decisions, as well as for finding a technical solution to eliminate these camera vulnerabilities, or selecting and testing another camera for our mobile technology platform. The results advance our theoretical knowledge and have a praxeological significance for the creation of a technological demonstrator and subsequently a prototype of a smart mobile airport security system. Institutions responsible for the protection of state borders, the fight against illegal migration, smuggling of goods, etc. are also interested in mobile security solutions. | 6,104 | 2020-09-30T00:00:00.000 | [
"Computer Science"
] |
Probability Analysis to Improve the Confidence in Profiling Accuracy
Performance profiling for the system is necessary and has already been widely supported by hardware performance counters (HPC). HPC is based on the registers to count the number of events in a time interval and uses system interruption to read the number from registers to a recording file. 'e profiled result approximates the actual running states and is not accurate since the profiling technique uses sampling to capture the states. We do not know the actual running states before, which makes the validation on profiling results complex. Jianwei YinSome experiments-based analysis compared the running results of benchmarks running on different systems to improve the confidence of the profiling technique. But they have not explained why the sampling technique can represent the actual running states. We use the probability theory to prove that the expectation value of events profiled is an unbiased estimation of the actual states, and its variance is small enough. For knowing the actual running states, we design a simulation to generate the running states and get the profiled results. We refer to the applications running on production data centers to choose the parameters for our simulation settings. Comparing the actual running states and the profiled results shows they are similar, which proves our probability analysis is correct and improves our confidence in profiling accuracy.
Introduction
In data centers, performance is critical to improve the quality of service [1] and save costs [2]. Multiple tasks controlled by the operating system share computation resources at the same time to improve user experience and resource utilization. e whole system's performance representing the combination of multiple tasks is not enough for analysis of each task's performance. Many applications need to know the running states of specific tasks. ese applications include anomaly detection on data centers [3,4], compiler optimization using method stacks [5,6], and hot spots detections [7,8].
Modern processors have hardware support to monitor system performance. Hardware Performance Counters (HPC) [9] are register-based counters to count the number of events in a time interval. With the help of interruption, HPC can output the counted number to a recording file. For profiling each task's performance, only one extra information is in need-the instruction address. e instruction address indicates which task the processor is working for at the moment of interruption. e profiling technique then treats the counted events in the last time interval as all caused by this task indicated by the instruction address. It is not accurate to use the instant instruction address to represent the running states of a long sustaining time interval. But it has already been used to profile the performance of tasks.
Many widely used profiling tools have already adopted this approximation method, like PAPI [10,11], perf [12], and VTune [13]. Profiling accuracy attracts research attentions. e experiment-based evaluation compares the profiling results across multiple system architectures to improve the confidence of the profiling technique [14,15]. And CPU simulator gives detailed information, which makes the comparison more direct [16]. But these researches utilized simple benchmarks to check the accuracy.
is kind of validations cannot deduce other workloads' conditions since the mechanism lacks proof and analysis-they have not explained why the sampling technique can represent the actual running states.
In this paper, we show the mechanism of the profiling technique with the help of probability theory. We model the profiling process with two main elements: the running granularity of a task and the sampling interval. We classify all possible conditions into three classes according to the rate between running granularity and sampling interval. For a constant rate, the sampling process is a kind of Bernoulli experiment [17]. We prove no matter what the rate value is, the expectation value is unbiased to the actual value, including the condition with a mixture of rate values that would still keep the unbiased property. And the variance is related to the number of samples, which is small enough and usually smaller than 0.25.
We further use the simulation experiments to validate our proof. e implementation of simulation includes the generation of actual running states and the sampling process. e settings of the simulation follow the characteristics of the workloads running on live production data centers. We simulate single tasks and mixed tasks running with multiple running granularities and under multiple resource utilization levels. All of these experiments show that the expectation value is an unbiased estimation. e variance is also included into consideration, whose effect does not influence the unbiased property.
We organize our paper as follows. Section 2 introduces the background of the profiling technique. We propose our analysis model in Section 3. Section 4 designs the simulation model. Section 5 shows the simulation results including the simulation's prerequisite. Section 6 reviews the related work. And we conclude in Section 7.
Profiling Technique
In this section, we introduce the profiling technique used for recording the running states of clusters. For example, Figure 1 is a profiled result of the Windows operating system. e green part at the bottom shows the value changes of CPU utilization in this 60-second observing window. e blue line shows the rate of current operating frequency to the highest frequency. ese lines link the samples of every second. And each sample represents the averaged CPU utilization of the last 1-second time interval.
HPC Profiles.
Hardware performance counters consist of two components-event detectors and event counters [18]. Users can configure performance event detectors to detect performance events as cache misses, cycles consumed, or branch mispredictions. Often, event detectors have an event mask filed that allows further qualifications of the event. According to the processors' privilege mode, for example, HPC can collect the kernel occurred events with the administrator mode. e event counter would increase itself by one if this event happened once until a system interruption happened.
It outputs its historical value and its value can be reset to zero or not according to whether it is on accumulative counting mode. e condition to cause this kind of system interruption can be separated into two types: (i) Time-based sampling is implemented through interrupting tasks' execution at regular time intervals and recording the program counters. is approach is often used to show the relationships between profiled events to time dimensions. (ii) Event-based sampling is implemented through interrupting after a specific number of performance events-when the number of events that happened reaches a threshold. Users can specify the threshold events.
e hardware performance counter method has distinct advantages. First, it profiles the system from the hardware level without any intrusion to applications, making applications and operating systems remain largely unmodified. Second, every modern processor has the support of performance counters. is method is a general solution. ird, this method profiles system on the fly as the applications executing to save the effort to reproduce the workloads, since some executions are prohibitively complex to be simulated or reproduced.
ough hardware performance counters reveal lots of information from the system view, this information mainly exposes the system's overall states, not a specific task's behaviours.
Task-Oriented Profiles.
For profiling the running states of tasks, only one extra information is in need to be added.
e information is the instruction address coming from context information [19,20]. We would set periodic events to trigger the sampling, like for every 100 cache misses. At the sampling moment, the recorded sample contains the content of performance counters and instruction address. en profiling technique uses this sample to represent the running of the last sampling interval. Figure 2 illustrates the profiling method. e upper bar is the actual running state, and the lower bar is the state captured by profiling. e profiled state is different from the actual running state. e upper blue block shows the actual running of task XX. e upper orange blocks show the actual running condition of task YY. And the vertical lines represent the sampling moments that are triggered by the event threshold or time limit. e first sample regarding the last sampling interval events was all caused by task YY. e second sample would treat the last sampling interval events caused by task XX shown in the lower bar, though YY was running in most of the second sampling interval tasks. e profiling result shown in this figure is that task XX and task YY both consumed 2 units of the resource. Our paper's task is an abstract presentation that can represent the threads, processes, programs, or applications according to the analysis granularity. If we intend to profile the performance of an application, then the task means the application. However, an application can be divided into multiple threads. All these threads are regarded as running for a single application-a single task.
Challenges.
Checking the profiling method's accuracy has many challenges from the complex environment and the limitations of profiling techniques. In the following, we list the changes from two aspects.
First, we do not have the standard answers to validate the profiling results. In the data center, there are tens of thousands of applications running on thousands of computers [21]. ese applications include online services that have high requirements on response time and off-line services that require high throughput. Sometimes the clusters can reach extremely high pressure, for example, the double 11 shopping festival. e complex environment makes every next moment different from the last one. We do not know the true proportions of the running applications or the true load of queries from users. Many production scenes appeared only once, which means these conditions could hardly be repeated.
Second, the profiling technique would unavoidably introduce overhead to the running system [22]. With the increasing sampling frequency, the overhead would increase, making it impossible to increase the sampling density too much to get detailed enough running states. And it is also impossible for current profiling techniques to separate the profiling workload from the original workload.
e experiments on benchmarks only prove some events' correctness under certain workloads, and these experiment-based researches have not covered all scenarios. An explanation of why profiling can be trusted would improve our confidence when profiling the system that has not been covered.
Analysis Model
3.1. Application Scenario. A representative scene using task-oriented profiling is the hot spot detection. Taking the hot methods detection as an example, it targets finding out the top hot methods that consumed the most resources (like CPU cycles) for further performance optimizations. Not every method can catch enough attention to be optimized further since there are too many methods running on a live environment to be optimized one by one. us for profiling how many CPU cycles are consumed by a method, we can set a sampling-based method to profile the running of methods.
For example, we set a sample of 0.1 seconds, which means every 0.1 seconds to interrupt the system running and record the current instruction address.
is instruction address indicates the running method, for example, is "Sort()," and the number of cycles consumed is 200 million in this sampling interval. en we count that the "Sort()" method consumed 200 million cycles. is interruption on the system is repeated to get an overview of the CPU cycle's consumptions of methods.
Model Components.
We model the profiling process as two major elements to help us do further analysis. e main elements that need to be considered include the following.
(i) Running granularity: e averaged scheduling time of a task running continually until being switched out. Running granularity would be influenced by many factors like the property of this task, our observing level, system environment, etc. e length of a color block shown in Figure 3 is called the running granularity. e running granularity does not need to be a constant value. (ii) Sampling interval: e distance between the last sampling to current sampling. Figure 3 shows an example. If the interruption is event-based rather than time-based, and the number of events is not proportional to time, the sampling interval's length would look nonuniform from the time dimension. But from the corresponding event dimension, it is still of uniform intervals.
In the following analysis, the base event is to denote the event dimension that causes the sampling interruption. For example, if it is time-based sampling, then the base event is time, and if it is CPU cycles based sampling (e.g., interrupt system every 250 million cycles), then the base event is CPU cycles.
e number of base events that happened in a sampling interval is a constant value without variance. We call the constant number of events that happened in a sampling interval a unit of events.
About the nonbase events, the numbers of these events collected by samples may not be as steady as the base event.
e number of nonbase events in a sampling interval would be different from the other sampling intervals. eir estimation variance would be a little higher than base events. We include the considerations on nonbase events by introducing variance to the constant unit of events when doing experiment. To avoid introducing extra variable considerations into probability model, we first model our analysis focusing on base event, which can be further extended into nonbase events by adding extra considerations on the variance of the unit of events.
ree Classes of Conditions.
e accuracy of estimations on the base event would be mainly influenced by the tasks switches reflected by the rate between running granularity and sampling interval. When the sampling interval becomes smaller, and the running granularity keeps the same, the sampling's accuracy would increase, and the error bound would be smaller. Assuming an extreme condition that the sampling interval equals every clock cycle, the profiled result reflects the real running state accurately without any approximation.
We utilize the rate between the running granularity and sampling rate to define all possible conditions. We define a variable R to denote the rate as
R �
Running Granularity Sampling Interval . (1) According to the value of R, there are three kinds of conditions as shown in Figure 4. ey are (a) R � 1, (b) R < 1, (c) R > 1.
(i) Figure 4(a) represents the R � 1 condition that the sampling interval and the running granularity are the same. (ii) Figure 4(b) represents the R < 1 condition that the running granularity is smaller than the sampling interval, which means it is possible that the tasks already have been switched more than once within one sampling interval. (iii) Figure 4(c) represents the R > 1 condition that the running granularity is larger than sampling interval.
We use these three kinds of conditions to help with our further analysis.
With a specific constant R value, the sampling process is a kind of Bernoulli experiment whose results would follow a binomial distribution.
e Bernoulli experiment means running a task would be captured by a sample or would not be repeated independently. We first use cases with representative R values to illustrate the calculation of the profiling distribution's expectation value, conclude them with a general representation method, and show its corresponding variance calculation method.
(1) For the first condition that R � 1, the sampling interval and the running granularity are of the same length. No matter where the sampling starts, one of two adjacent samples would capture this task and regard it caused by one unit of base events-one unit of base events means the constant number of base events that happened in a sampling interval. is is an accurate estimation without errors. Additionally, when the R value is integer, like 2 or 3, the sampling result would keep the same condition as the R � 1 and give an accurate estimation. (2) For the second condition that R < 1, we denote the time that the sample is captured as e i value means this is the i th sample, and the profiling starts from t 0 with a period of T p . Every T p would occur an interruption to get the sample. ere is an assumption on the t 0 . We regard the time starting to profile (t 0 ) as randomly chosen-t 0 is independent of T p or other factors. We assume r proportion (r < 1, e.g., 30%) of sampling interval is working for a task. e probability of being captured by a sampling point equals the running proportion of this task as r(P(captured) � 30%) in this sampling interval since the sampling point is independent of running this task. And the probability that this task is missed (not captured by the sampling is process is repeated, and we get a bunch of profiled samples. When we denote a unit of events as #(events), the expectation value of events caused by this task in an interval can be deduced as E(events) � P(captured) · #(events) + P(missed) · 0 � r · #(events).
(3)
We can find that the expectation value equals the real running proportion.
(3) For the third condition that R > 1, we first analyze the case when the task's running granularity is less than 2 times of sampling interval and bigger than 1 time of sampling interval denoted as r times of sampling interval length. ere are two possible conditions. One is shown in Figure 4(c) that this task appears in three intervals and is captured by two samples. Another one is shown in Figure 5 that this task appears in two intervals and only is captured by one sample. e probability of the first condition captured by two samples is P(two) � r − 1, and the probability of the second condition captured by a single sample is the left probability P(one) � 1 − P(two) � 2 − r. e expectation value of R > 1 can be combined from these two probabilities as where the #(events) represents a unit of events. is condition would come to unbiased estimation too.
Piecewise Binomial Distribution.
We conclude this deduction process to be a more general representation. e running granularity is r times of sampling interval. All possible sampling conditions are separated into two classes-when the running is captured by int(r) samples and captured by int(r) + 1 samples. e probabilities of being captured by int(r) + 1 samples and int(r) samples equal to P(int(r) + 1) � MOD(r, 1) and P(int(r)) � 1− MOD(r, 1). And the expectation is E � P(int(r) + 1) · (int(r) + 1) · +P(int(r)) · int(r) where the int(·) is a function to get the integer part of this element and the MOD(r, 1) is an operation to get the fractional part. For example, int(3.14) equals 3 and MOD(3.14, 1) equals 0.14.
is expectation value shows profiling method would get an unbiased expectation value about the running proportion.
We can find the sampling result distribution can be regarded as a binomial distribution under a specific constant R value. According to the binomial distribution, we get the variance (V) of sampling distribution under a specific r value as V � MOD(r, 1) · (1 − MOD(r, 1)).
is expectation value and variance value is about the distribution where samples are drawn. It is an ideal model, and enough number of samples can approach its distribution according to Chebyshev's theorem [23].
In actual running conditions, the running granularity would change, which makes the R value vary. We can regard the varying R value as the mixture of components with different R values. But no matter what kind of mixture proportion of these components, the expectations of components are all equal to r value. en the combined expectation is equal to r. is can be denoted as where the Pro i represents the proportion of i th component, the E i represents the expectation value of i th component. e property that the expectation value is unbiased still exists when considering varying R values.
e Number of Samples.
e number of samples would influence the approximation to the distribution. When the number of samples is n, then the mean value of these samples would keep the same to expectation value, and the variance of samples would be related with n: e maximum variance value appears as 0.25, when MOD(x, 1) � 0.5 and n � 1. e small variance can improve our confidence in the current profiling method. Taking the worst condition as an example, it still has good performance. e probability of misestimating the mean value by one more unit event is less than 0.022 8.
Simulation Model Design
We introduce how to implement our simulation process. Two parts need to be modelled: the actual running states and the profiling process.
It is a reflection of the condition that multiple tasks share the computing resources about the running state generation. For simplifying the resources, we will not detail computing resource into more specific types. We treat the resource such that it can only be occupied by one task at one moment, which is a simplified model to the true system. But model duplications to consider multiple types of resources are similar to the true system. For example, the multiprocessor CPU would run multiple processes simultaneously. At one moment, the CPU resource of the system can be shared by different tasks. But this true condition can be simulated by repeating this simplified model. A running model represents one processor resource that runs a single process at one moment. is simplification keeps the basic property of the system. If necessary, this model can be rebuilt into a complex system. For simulating the running states, we regard the sample sample sample Mathematical Problems in Engineering smallest resource amount allocated to run tasks as a resource unit (e.g., a CPU cycle). Several parameters need to be specified, including the work amount of tasks (the number of resource units needed), their corresponding running granularity, and their scheduling and sharing behaviour. Regarding the sampling method, the sampling interval is the key setting. By changing the sampling interval and running granularity, we simulate different R value conditions. e sampling interval would keep stable roughly with little variance. We add random noise into the sampling interval to approach the real profiling scene. e start time of the sampling t 0 can be generated randomly. We get the profiled samples based on the same true running condition, repeating the sampling with different randomly generated start time and different noise introduced. e statistics of profiled samples about the proportion of tasks are supposed to approach real task running proportion values.
Experimental Evaluation
is section shows the simulation results on various kinds of single task and mixed tasks, including the variance introduced to simulate nonbase events. We first introduce the simulation components and the experimental results are followed to prove the effectiveness of our model.
Simulation Components.
Our experiments are conducted on real running environments and the analysis data is collected by the system profiling tool. We call it a simulation experiment since the workload that is running on the system is simulated without actual functions and is under control.
ere are three components that need to be clarified, including data collection method, the control of workloads' running granularity, and the method to introduce the nonbase event variance.
Data Collection Method.
We use the "perf" to profile the system-the Linux kernel already contains this profiling tool. An application or a program serving for user requests consists of multiple methods, and these methods belong to their corresponding modules. e "perf" script would offer an automatic parsing function to map Instruction Pointer (IP) value to the corresponding method and module, record the hardware events, and index each sample with the sampling timestamp, CPU number, and event name. e dimensions collected by the "perf record" script related to our model and their meanings are shown in Table 1. We profiled the running states of these physical machines from CPU view-each recorded sample represents the state of a CPU in a sampling interval.
Running Granularity Scale.
We regard the continuing samples with the same module name as the condition that this module has not been switched out-its running granularity value is calculated as the sum of the continuing sampling interval. e distribution of sampling interval from the time dimension is shown in Figure 6. e distribution of running granularity deduced from the sampling results is shown in Figure 7. When the running granularity is smaller than the sampling interval, we treat it equal to the sampling interval, making running granularity here a little larger than its actual value. e running granularity is not always smaller than the sampling interval. It is about several times of sampling interval.
us the R value scale that we experiment with, like 1 to 10, is good enough to cover the conditions rather than a thousand or million scale.
Nonbase Event Variance.
For showing the variance of nonbase event is small when profiling and our model about the variance of nonbase event is reasonable, we analyze the variance of "cycle" event when the base event is "time." e sampling intervals in the collected data are not the same; thus, we scale the cycle event value by dividing the number of cycles by the sampling interval's length. For example, the distribution of profiled cycle event in one physical machine that ran for 231 modules in 5 seconds is shown in Figure 8. e cycles consumed in a sampling interval tend to be fully utilized or in idle state, but this characteristic does not make the variance large for each module. For each module, we filter out its corresponding cycle event samples and calculate its variance. For these 213 modules that appeared in our 5second observing window, not every module has a large number of samples. us we filter out 111 modules whose number of samples is larger than 10. We get the variances of these 111 modules shown in Figure 9. e mean value of the variances is 0.317 1. e variance we introduce to simulate nonbase events between 0 to 2 is in a reasonable range. Figure 6: e distribution of sampling intervals when profiling the data center. is is a histogram whose x-axis represents the time length of sampling interval and y-axis represents the number of samples belonging to a specific value range. 6 Mathematical Problems in Engineering
Unbiased Estimation.
In this section, we show that, under any R condition, the expected value of profiled samples is unbiased, and their variance is small. We first use multiple types of tasks respectively with different utilization levels and keep the sampling interval the same-which means different R values. And we also show the estimation of mixed tasks with different running granularity also has good performance.
Single Task.
We set 1 million units to simulate running states, and each unit can work for a task or in idle. We set the utilization of units as 80%, 50%, and 30% to combine with the running granularity as 30, 50, 100, 150, 200, and 280 separately to cover a total of 18 types of running states. en we profile these 18 types of running states by setting every 100 units to trigger interruption to get a sample-the sampling interval is 100, and repeat profiling on each running state 1000 times and get 1000 profiling results of each running state.
We do not introduce any extra variance first and randomly generate the running states for each unit according to the set running granularity. Taking the 80% utilization load level as an example as shown in Table 2, the mean value of utilization for these 1 million units approach to the set utilization value. But within each sampling interval, this task's utilization is not always 80%, as shown in Figure 10-we take 30 running granularity as an example to plot out the actual running states for each sampling interval. e profiling process is conducted on these 18 types of running states. We find profiling results have little variance and are unbiased to the expectation value-the detailed result of 80% is shown in Table 2. e mean value distribution of 1000 times profiling result on 30 running granularity and 80% utilization condition is shown in Figure 11 whose minimum value is 0.786 9 and maximum value is 0.809 2. e profiling results of the other conditions with 50% and 30% utilizations are shown in Table 3 whose expectation values approach the actual utilization value and variance values are small too.
But we may doubt the low variance of these 1000 profiling results caused by the large number of samples collected by each profiling process-reaching 10 thousand samples. Of course, a large number of samples would guarantee a low variance. e fact is that the variance is still small enough even when the number of samples is only 1. We reduce the number of samples to 1. For example, the variance of 80% utilization when running granularity is 30 equals 0.004. us we can believe that, without extra variance introduced, the estimation is unbiased, and its variance is small.
Mixed
Tasks. Except for a single task running on a system, it is more usual that multiple types of tasks (with different running granularity) share a system simultaneously. We simulate this condition by mixing tasks with a specific proportion. We show the result to make the system with 80% utilization composed by 30% task 1 with 30 running granularity, 20% task 2 with 50 running granularity, 20% task 3 with 150 running granularity, and 10% task 4 with 280 running granularity. We observe on 1 million units, and the sampling granularity is still 100. e generated running states show that each task's mean value is unbiased, as shown in Table 4 with small variance. We can find the mixed running still keeps the unbiased property. : e distribution of variance values of 208 modules running on a physical machine. In this histogram, each sample means a variance value of a module, whose x-axis represents the variance value and y-axis represents the number of modules in a specific value range. e variance of module is calculated by its corresponding profiled samples. Figure 7: e distribution of running granularity deduced from sampling results. is is a histogram whose x-axis represents the time length of running granularity and y-axis represents the number of samples belonging to a specific value range.
Variance Introduced to Simulate Nonbase Events.
e sampling interval and unit of events are constant values in the former experiments-having no variance. But the real condition would always have some variations.
us we explore the impact of the variance of sampling interval and unit of events in this section. Some nonbase events would not be as accurate as base events. us we introduce variance to sampling interval or unit of events to simulate the performance of nonbase events. e inaccuracy of the sampling interval would influence the specific number of the units of events. us the variance introduced to sampling interval or unit of events has the same effect to simulate.
We introduce the noise to a unit of events-the number of events counted by a sample varies by a noise value generated randomly from a normal distribution. We denote the normal distribution by N(mean, standard deviation) function. We profile the same running states of mixed tasks as the former section and introduce noise to a unit of events whose value is regarded as 100 before-since the sampling interval is set to 100.
e noise for each sample is drawn from the normal distributions N(0,0.5), N(0,1), and N(0,2), respectively, and the unit of events profiled for each sample is calculated by e mean and standard deviation (SD) value of 1000 times profiling-each profiling gets about 10 thousand samples-with noise introduced following three normal distributions, respectively, are shown in Table 5.
We also reduce the number of samples collected by each profiling process. When the number of samples is reduced to 1 thousand, it is shown in Table 6. e estimation is still unbiased, and the variance is influenced by injected noise but not too much to influence the mean value. e different actual running states cause the differences between task 1 and task 4 when n � 500 and noise following N(0,1) distribution-the real proportions of task 1 and task 4 are 0.343 2 and 0.556 0. is means they are still unbiased estimations on the real proportions. ere are 1000 samples. One profiling on system means one sample here. Any profiled result is around the correct answer with good performance.
Related Work
Except for the hardware-based performance profiling, there are another two representative methods. One is an intrusive method [19,24,25] that needs to modify the application source code to add instrumentation code for collecting data. is method requires the authority to access the source code, rebuild the application source code, and redeploy this version into the system. ese requirements are impractical. Moreover, these intrusive methods can disturb the application's behaviour, bringing other questions about the collected data's validity.
Another one is the simulator-based method [26,27] using the processor simulator that models the real processor's architecture. It collects processor performance data by using the simulator to execute the application. is method can yield detailed data on a processor like the pipeline stalls and cache line behaviours. However, not every processor would have its corresponding simulator that is provided by its manufacturers. e simulator would also be tens of times slower than running on real processors, making performance profiling costly. e task granularity profiling is useful in code profiling and hot execution path detection [28]. Identifying program hot spots can support runtime optimization [29,30]. e application anomaly [4] or stragglers detection [31,32] also needs the information from the task-level.
For more accuracy to count the events into specific tasks, there are instruction-oriented profiling techniques [33]. e profiling interruption is triggered by an instruction related dimension. A detailed record of interesting events and pipeline stage latencies in the out-of-order processor is collected. Trace-based profiling [34,35] has a similar design to follow a running pipeline to record the running states. But they are not useful in improving the confidence in the hardware counter accuracy.
Many pieces of literature analyze accuracy from the probability theory. Chen [36] proposed the ProbPP method for analyzing the probabilities on the execution paths of the multithreaded programs. Yan and Ling [37] used the probability model on the memory level parallel analysis to estimate the maximum number of cache misses. But they did not use the probability model to prove the accuracy of hardware-based profiling technique.
Conclusion
In this paper, we analyze the hardware-based profiling technique's mechanism using the probability theory and design an analysis model to simulate the profiling process. e setting of the simulated model follows the characteristics of workloads running in a live environment. e simulation results validate our probability deduction result and show that the expectation value has nonbiased property, and the variance is small. It is expected that this work can improve confidence in the profiling accuracy and broaden the relevant research directions.
Data Availability e simulated running state data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
e authors declare that they have no conflicts of interest. Mathematical Problems in Engineering 9 | 8,199.8 | 2021-11-30T00:00:00.000 | [
"Computer Science"
] |
Reflectance Estimation from Multispectral Linescan Acquisitions under Varying Illumination—Application to Outdoor Weed Identification †
To reduce the amount of herbicides used to eradicate weeds and ensure crop yields, precision spraying can effectively detect and locate weeds in the field thanks to imaging systems. Because weeds are visually similar to crops, color information is not sufficient for effectively detecting them. Multispectral cameras provide radiance images with a high spectral resolution, thus the ability to investigate vegetated surfaces in several narrow spectral bands. Spectral reflectance has to be estimated in order to make weed detection robust against illumination variation. However, this is a challenge when the image is assembled from successive frames that are acquired under varying illumination conditions. In this study, we present an original image formation model that considers illumination variation during radiance image acquisition with a linescan camera. From this model, we deduce a new reflectance estimation method that takes illumination at the frame level into account. We experimentally show that our method is more robust against illumination variation than state-of-the-art methods. We also show that the reflectance features based on our method are more discriminant for outdoor weed detection and identification.
Introduction
Nowadays, one of the biggest topics in precision farming is to increase yield production while reducing the quantity of chemicals. In order to optimize the application of herbicides in crop fields, less toxic and expensive weed control alternatives can be considered due to recent advances in imaging devices. During the last decade, sophisticated multispectral sensors have been manufactured and deployed in crop fields, leading to weed detection [1][2][3].
Multispectral cameras collect data over a wide spectral range and they provide the ability to investigate the spectral responses of soils and vegetated surfaces in narrow spectral bands. Two main categories of devices can be distinguished in multispectral image acquisition. "Snapshot" (multi-sensor or filter array-based) devices build the image from a single shot [4]. Although this technology provides multispectral images at a video frame rate, the few acquired channels and low spatial resolution may not be sufficient for fully exploring the vegetation spectral signatures. "Multishot" (tunable filter or illuminationbased, push-broom, and spatio-spectral linescan) devices build the image from several and successive frame acquisitions [5][6][7]. Despite being restricted to still scenes, they provide images with a high spectral and spatial resolution. We use a multishot camera, called the "Snapscan" to acquire outdoor multispectral radiance images of plant parcels in a greenhouse under skylight [8]. From this radiance information, reflectance is estimated as an illumination-invariant spectral signature of each species. Several methods have been proposed for computing reflectance thanks to prior knowledge regarding cameras or illumination conditions [9][10][11]. In field conditions, the typical methods first estimate the illumination by including a reference device (a white diffuser or a color-checker chart) in the scene [2,[12][13][14]. Subsequently, reflectance is estimated at each pixel p by the channel-wise division of the value of the radiance image at p by the pixel values that characterize the white diffuser or the color-checker white patch. In [15], an extension to the multispectral domain of four algorithms traditionally applied to RGB images is proposed for estimating the illumination. In [16], a Bragg-grating-based multispectral camera acquires outdoor radiance images and reflectance is estimated from two white diffusers, one along the image bottom border and another fully visible one. A coarse reflectance is computed first, and then rescaled using illumination-based scaling factors. Finally, the resulting reflectance image is normalized channel-wise by the average reflectance value that is computed over pixels of the second white diffuser that is present in the scene. In [17], a multispectral camera is used in conjunction with a skyward pointing spectrometer to estimate the reflectance from the acquired scene radiance.
These methods require additional devices and knowledge regarding the spectral sensitivity functions (SSFs) of the sensor filters. They also often assume constant incident illumination throughout a few seconds. However, in outdoor conditions, illumination may vary significantly during the successive frame acquisitions (scans) that last for several seconds.
In this paper, we propose a reflectance estimation method that is robust to illumination variations during the multispectral image acquisition that was performed by the Snapscan camera. In Section 2, we provide details regarding multispectral radiance image acquisition with this device and propose an original model for such image formation. In Section 3, this model is first used to study reflectance estimation under constant illumination. We also show that, when outdoor illumination varies during the frame acquisitions, the assumption about spatio-spectral correlation does not hold. Based on this model, we then propose a new method for estimating the scene reflectance from a multispectral radiance image acquired in uncontrolled and varying illumination conditions (see Section 4). Section 5 presents an experimental evaluation of the proposed reflectance estimation, and Section 6 shows the results of weed/crop segmentation using the estimated reflectance.
Multispectral Radiance Image Acquisition by Snapscan Camera
In this section, we first detail how the Snapscan camera achieves radiance measurement and frame acquisition. Subsequently, we explain how a multispectral image is obtained from the successively acquired frames. We propose an original multispectral image formation model that handles how illumination is associated to both the considered band and pixel because the radiance that is associated to a given spectral band at a pixel is measured in a frame that is acquired at a specific time. From this new model, we show that the spatio-spectral correlation assumptions do not hold when illumination varies during the frame acquisitions.
Radiance Measurement
The Snapscan is a multispectral camera manufactured by IMEC that embeds a single matrix sensor that is covered by a series of narrow stripes of Fabry-Perot integrated filters. It contains B = 192 optical filters whose central wavelengths range from λ 0 = 475.1 nm to λ B−1 = 901.7 nm with a variable center step (from 0.5 nm to 5 nm). Specifically, each filter of index b ∈ [0, B − 1] is associated with five adjacent rows of 2048 pixels that form a filter stripe, and it samples a band from the visible or near infra-red spectral domain according to its SSF T b (λ) with a full width at half maximum between 2 nm and 10 nm.
The Snapscan camera acquires a sequence of frames to provide a multispectral image. During frame acquisitions, the object and camera both remain static while the sensor moves and illumination may change. Therefore, the measurement of the radiance that is reflected by a given lambertian surface element s of the scene varies according to the frame acquisition time t, although s is projected at a fixed point q of the image plane. Let us denote, as E t (λ) ∈ [0, 1], the relative spectral power distribution (RSPD) of the illumination at t and assume that it homogeneously illuminates all of the surface elements of the scene. The radiance that is reflected by s and refracted by the camera lens projects onto the image plane at q as a stimulus L t,q (λ): where R q (λ) ∈ [0, 1] is the spectral reflectance of the surface element s that is observed by q, and A q (λ) ∈ [0, 1] is the optical attenuation of the camera lens at q. All of these functions depend on the wavelength λ. The sensor moves forward on the image plane according to the direction perpendicular to the filter stripes (see Figure 1a). Between two successive frame acquisitions, it moves by a constant step v = 5 (in pixels) that is equal to the number of rows in each stripe. Therefore, the radiance that is measured at q is filtered by a different Fabry-Perot filter of index b t,q (0 ≤ b t,q < B) at each acquisition time t. The radiance at q is fully sampled over N frame acquisitions, provided that each of them measures the radiance there, i.e., N ≥ B. Let the coordinates of point q be (x q , y q ) C in the camera 2D coordinate system (O, x, y) whose origin O corresponds to the intersection between the optical axis and image plane. The unit vectors of x and y are given by the photo-sensitive element size (i.e., axis units match with pixels), and y is oriented opposite to the sensor movement. At a given point q, the filter index b t,q can then be expressed as: where y 0 is the coordinate along y of the first filter row at first acquisition time t = 0. Note that the light stimulus L t,q is only associated to a filter at a given point q when t 0 q ≤ t < t B q . The lower bound t 0 q = (y 0 − y q )/v is the acquisition time at which the first optical filter of the sensor observes L t,q . The upper bound t B q = (y 0 − y q )/v + B is the time at which all of the sensor filters have observed L t,q .
Besides , at a given time t, the coordinate y q of point q that is associated to a photosensitive element of the sensor satisfies: since 0 ≤ b t,q < B. Given these restrictions, the radiance S t,q that is then measured at q by the sensor at acquisition time t is expressed as: where Q is the quantization function according to the camera bit depth, τ is the integration time of the frames, and Ω is the working spectral domain. Note that τ is set to the highest possible value that provides no saturated pixel.
Frame Acquisition
The radiance that is measured at q is stored by the camera as a pixel value f t,q = S t,q in frame f t (see Figure 1b). We define the coordinate system (O , x , y ) attached to the sensor, such that origin O is the first (top-left) photo-sensitive element location, axis y corresponds to y, and x is parallel to x, in order to compute the coordinates of q relative to the frame. In this frame system, the coordinates of q are (x q , y q ) F = (x q , y q − y 0 + t · v) F . Note that Equation (3) allows us to check that 0 ≤ y q < B · v.
Snapscan camera Conversely , any given pixel p(x p , y p ) F of a frame f t is mapped to the coordinates in the camera coordinate system as: at which the stimulus L t,p (λ) of a surface element radiance is filtered by the filter of index b p = y p /v . From this point of view, each frame pixel value is, therefore, also expressed as: Before the frame acquisitions, the Snapscan uses its internal shutter to acquire a dark frame f dark whose values are subtracted pixel-wise from the acquired frames. Therefore, we assume that the pixel value that is expressed by Equation (6) is free from thermal noise.
Let us also point out that, at two (e.g., successive) acquisition times t 1 and t 2 , the sensor is at different locations. Therefore it acquires the values f t 1 ,p and f t 2 ,p from the stimuli L t 1 ,p and L t 2 ,p of two different surface elements at a given pixel p whose coordinate y p in the camera system is time-dependent (see Equation (5)). Besides, the stimuli L t 1 ,p and L t 2 ,p are filtered by the same filter whose index only depends on the pixel coordinate y p in the frame system. Equations (4) and (6) model the radiance that is measured at a given point in the image plane and stored at a given pixel of a frame, respectively. Both of the equations take account of illumination variation during the frame sequence acquisition, but differently take the sensor movement into account. Indeed, the filter index changes at a given point of the image plane during the frame acquisition (see Equation (4)), whereas the observed surface element changes at a given pixel in the successive frames (see Equation (6)).
Stripe Assembly
We now determine the first and last acquisition times of the frame sequence that is required to capture an object of interest whose projection points on the image plane are bounded along the y axis by q a (x q a , y q a ) C and q m (x q m , y q m ) C , with y q a > y q m . Given the initial coordinate y 0 of the sensor along y, we can compute the first and last frame acquisition times t 0 q a and t B−1 q m , so that the measured radiances at the points between q a and q m are consecutively filtered by the B sensor filters (see the top part of Figure 2). The acquisition of the multispectral image from the frame sequence { f t } t B−1 qm t=t 0 qa takes account of the spatial and spectral organizations of each frame. A frame f t is spatially organized as juxtaposed stripes of v adjacent pixel rows. A stripe f b t , b = 0, . . . , B − 1, of v adjacent pixel rows contains the spectral information of the scene radiance that is filtered according to the SSF T b (λ) of filter b centered at wavelength λ b . All of the stripes that are associated with filter b in the acquired frames are stacked by the assembly function to provide a stripe assembly defined as: The size of each stripe assembly is 2048 pixels in width and N · v pixels in height, where N = t B−1 q m − t 0 q a /∆ + 1 is the number of acquired frames and ∆ is the frame acquisition period.
To form the multispectral image I (B) = {I b } B−1 b=0 of the object of interest, only the scene part that is common to all stripe assemblies is considered by the camera (see the bottom part of Figure 2). Specifically, the retained stripes in the b-th assembly are acquired between t b q a and t b q m to form each channel I b : The multispectral image I (B) has its own coordinate system. For convenience, in the sequel, we denote a pixel as p(x, y) in this system, since the camera and frame coordinate systems are not used any longer.
Formation Model of a Multispectral Image Acquired by Snapscan Camera
We can now infer an image formation model for multishot linescan cameras, such as the Snapscan. At any pixel p, the radiance value (8)). It results from the light stimulus L t b p ,p that was filtered according to T b (whose index dependence upon p is dropped by stripe assembly step), and is therefore defined from Equations (1) and (6), as: The term E t b p (λ) shown in Equation (9) points out that illumination is associated to both a channel index and a pixel. These dependencies may weaken the spatio-spectral correlation assumptions of the measured scene radiance.
Spectral correlation relies on the assumption that the SSFs that are associated to adjacent spectral channels strongly overlap. Thus, radiance measures at a given pixel in these channels should be very similar (or correlated). Let us consider the radiance values in two channels b 1 and b 2 at a given pixel p. Even if the SSFs T b 1 (λ) and T b 2 (λ) strongly overlap (and are equal in the extreme case), the illumination conditions at t b 1 p . Spatial correlation relies on the assumption that the reflectance across locally close surface elements of a scene does (almost) not change. Thus, under the same illumination, the radiance measures at their associated pixels within a channel are correlated. Let us consider two pixels, p 1 (x p 1 , y p 1 ) and p 2 (x p 2 , y p 2 ), which observe surface elements of a scene with the same reflectance R p 1 (λ) = R p 2 (λ) for all λ ∈ Ω. If |y p 1 − y p 2 | ≥ v, then the radiances at p 1 and p 2 are acquired at different times t b p 1 and t b p 2 associated to different illumination conditions E t b p1 and E t b p2 , hence I b p 1 = I b p 2 . Therefore, the spatio-spectral correlation assumption does not hold in the image formation model of the Snapscan camera when illumination varies.
Reflectance Estimation with a White Diffuser under Constant Illumination
This short section introduces how to estimate reflectance by a classical (white diffuserbased) method and how the result should be post-processed to ensure its consistency.
Reflectance Estimation
In order to estimate spectral reflectance from radiance images that were acquired under an illumination that is almost constant over time, one classically uses the image I (B) [WD] of a white diffuser acquired in full field beforehand and assumes that: (i) The illumination is spatially uniform and it does not vary during the frame acquisitions, and p ∈ I (B) , and Equation (9) becomes: Note that the quantization function Q is omitted here, since the different terms are considered as being already quantized.
(ii) Each of the Fabry-Perot filters has an ideal SSF 0 otherwise, such that Equation (10) becomes: Reflectance is then derived for any pixel p that is associated to a spectral band centered at λ b as: The white diffuser is supposed to be perfectly diffuse and reflect the incident light with a constant diffuse reflection factor ρ wd . Hence, for I (B) [WD], we can write: where τ wd is the frame integration time of I (B) [WD]. Plugging Equation (13) into (12) yields the reflectance image that is estimated from a B-channel radiance image I (B) : This reflectance estimation model implicitly compensates the vignetting effect, since the white diffuser and object (scene of interest) occupy the same (full) field of view. Accordingly, I b p and I b p [WD] are affected by the same optical attenuation whose effect vanishes after division.
The estimated B-channel reflectance imageR (B) should then undergo two postprocessing steps: spectral correction and negative value removal.
Spectral Correction
Each of the Snapscan Fabry-Perot filters is designed to sample a specific spectral band from the spectrum according to its SSF T b (λ). However, because of the SSFs and optical properties of some filters (angular dependence [18], high-energy harmonics), several spectral bands are redundant, which limits the accuracy of the spectral imaging system. This leads to redundancy in spectral bands and introduces spectral information bias. Therefore, the reflectance image with B = 192 spectral channels is spectrally corrected and only K = 141 channels are kept in practice.
The spectral correction ofR (B) provides a spectrally corrected K-channel reflectance imageR (K) that is expressed at each pixel p as: where M is the sparse K × B correction matrix that is provided by the calibration file of our Snapscan camera. The linear combinations of the channel values ofR for the bands (referred to as "virtual" bands by IMEC) that are associated to the image channels, but the spectral working domain Ω = [475.1 nm, 901.7 nm] is unchanged.
Negative Value Removal
The acquired radiance image contains negative values due to dark frame subtraction, when the value of a dark frame pixel is higher than the measured radiance at this pixel. This generally occurs in low-dynamics channels, where the central wavelengths are in the range [475.1 nm, 560.4 nm] (before spectral correction). These negative values may lay on vegetation pixels and corrupt reflectance estimation at these pixels. Because we intend to classify vegetation pixels, this could lead to unexpected prediction errors. Negative values also occur-for even more pixels-in the spectrally-corrected reflectance imageR (K) (see Equation (15)), because the correction matrix M contains negative coefficients. Negative values have no physical meaning and they must be discarded. Because our images mostly contain smooth textures (vegetation, reference panels, soil), we consider that, unlike radiance, reflectance values are highly correlated over close surface elements.
Thus, we propose correcting negative values in imageR (K) by conditionally using a 3 × 3 median filter, as:R whereR k ref,p is the final reflectance value at pixel p for channel k. Because we consider the reflectance that is estimated by this model (Equations (14)-(16)) as a reference, it is denoted asR (K) ref .
Outdoor Reflectance Estimation with Reference Devices in the Scene
Because illumination varies during the acquisitions of outdoor scene images, the reflectance estimation method that is described by Equation (14) is not adapted to linescan cameras, such as the Snapscan. In such a case, one solution is to use several reference devices [16].
As a first reference device, we use a white diffuser tile mounted on the acquisition system, so that the sensor vertically observes a portion of it (see Figure 3a). Therefore, the pixel subset W D contains (about 10%) right border pixels that represent the white diffuser, as shown in Figure 3b. Because W D spans all the image rows, we further extract a small white square W S that represents a sample of this reference device. Each acquired image also contains a GretagMacbeth ™ ColorChecker that is principally used to assess the performances that are reached by reflectance estimation methods. The pixel subset W P representing the ColorChecker white patch is used as a second reference device by the double white diffuser (dwd) method [16] that we have adapted to our Snapscan acquisitions, as described in Appendix A.
Although the vignetting effect only depends on the intrinsic camera properties, this method corrects it in each acquired image. In Section 4.1, we propose performing this correction by the analysis of the white diffuser image I (B) [WD]. Subsequently, we present state-of-the art estimation reflectance methods that only require one reference device in the scene, but assume that the illumination is constant during the frame acquisition. We finally propose a single-reference method to estimate reflectance in the case of varying illumination.
Vignetting Correction
The dwd method acquires a full-field white diffuser image before each scene image acquisition in order to correct the vignetting effect [16]. However, this procedure can be cumbersome, since it requires an external intervention in order to place/remove the full-field white diffuser. Other methods that are presented in the following only require correcting it only once. Because we consider that vignetting only depends on the intrinsic geometric properties of the camera, we propose correcting it thanks to the analysis of a single full-field white diffuser image I (B) [WD] acquired in a laboratory under controlled illumination conditions. The vignetting effect refers to a loss in the intensity values from the image center to its borders due to the geometry of the sensor optics. To highlight how the vignetting effect would affect radiance measurements, let us rewrite Equation (9) under the Dirac SSF assumption as: In order to compensate for the spatial variation of A p (λ b ), we compute a correction factor at p, because it requires no knowledge regarding the optical device behavior [19]. Being deduced from the full-field white diffuser image I (B) [WD], the correction factor is channel-wise and pixel-wise computed as: where I b [WD] is the median value of the m pixels (m = 11 in our experiments) with the highest values over I b p [WD], which discards saturated or defective pixel values. The correction factors are stored in a B-channel multispectral image, denoted as C.
Because C is deduced from a single white diffuser image, it would be corrupted by noise (even after thermal noise removal during the frame acquisitions). Thus, we propose to directly denoise C by convolving each of its channels C b with an 11 × 11 averaging filter H: The vignetting effect in the B-channel radiance image I (B) is corrected channel-wise and pixel-wise using the smoothed correction factors: where I b p andĨ b p are the intensity values before and after vignetting correction. This procedure should reduce noise while preserving image textures. We assume that the attenuation is spatially uniform after vignetting correction (i.e., A p (λ b ) ·C b p = α b ∈ R for any given channel index b and pixel p), such that each value of the vignetting-free radiance image is expressed from Equation (17) as:
Reflectance Estimation with One Reference Device under Constant Illumination
The illumination E t b p (λ b ) that is associated to p can be determined using the radiance measured at a white diffuser pixel p WD ∈ W D. To determine illumination thanks to a single white diffuser as reference device, the methods in the literature often assume that illumination is constant, i.e., (21) then becomes , since the white diffuser has a homogeneous diffuse reflection R p WD (λ b ) = ρ wd (95% in our case). The reflectance at p is then deduced fromĨ b p andĨ b p WD as: To be robust against spatial noise, the white-average (wa) method [2,20] averages all of the values over the white diffuser pixel subset W S (see Figure 3b) and estimates the reflectance at each image pixel as: where | · | is the set cardinal. Similarly, the max-spectral (ms) method [15] assumes that the pixel with maximum value within each channel can be considered to be a white diffuser pixel for estimating the illumination. While ignoring the diffuse reflection factor, reflectance is estimated at each pixel in each channel by the ms method, as: where X contains all of the image pixels, except W D, and those of the ColorChecker. The wa and ms-based B-channel reflectance images undergo spectral correction and negative value removal (see Equations (15) and (16)) to provide the final K-channel reflectance imagesR
Reflectance Estimation with One Reference Device under Varying Illumination
In varying illumination conditions, the Snapscan acquires each row at a given time, hence under a specific illumination (see Section 2.4). Hence, reflectance can no longer be estimated, as in Equation (14). Instead, we propose determining the illumination that is associated to each row of the vignetting-free imageĨ (B) from the white diffuser pixel set W D [21]. The underlying assumption is that illumination is spatially uniform over each row at both the white diffuser and scene pixels (that may be not verified in the case of shadows). Based on this row uniformity assumption for illumination, we estimate reflectance fromĨ (B) in a row-wise manner, as follows. At pixel p with spatial coordinates x p and y p , Equation (21) can be rewritten as: To determine the illumination E t b yp (λ b ) that is associated to the row of p for channel index b, we use a white diffuser pixel r WD ∈ W D located on the same row as p. At r WD , the reflectance is equal to the white diffuser reflection factor ρ wd , and Equation (21) provides the vignetting-free radiance as: Because p and r WD are located on the same row, t b according to the assumption regarding the spatial uniformity over each row. Therefore, Equation (26) can be rewritten as: which can be considered to be an estimation of the illumination that is associated to pixel p.
For robustness sake, we propose computing it from the median valueĨ b W D,y p of the m highest pixel values that represent the white diffuser subset W D in y p , rather than from a single valueĨ b r WD . Plugging Equation (27) in (25) yields our row-wise (rw) reflectance estimation at pixel p for channel index b: In practice, setting m = 11 pixels is a good compromise for accurately estimating the illumination for each row and each channel.
Experiments about Outdoor Reflectance Estimation
We now present the experimental setup and metrics that were used to objectively evaluate the estimated reflectance. The accuracy results are obtained and discussed for the previously described estimation methods as well as for the extra training-based method described in the present section.
Experimental Setup
An acquisition campaign that was conducted in a greenhouse under skylight (see Figure 4a) provided 109 radiance images of 2048 × 2048 pixels × 192 channels of 10-bit depth. Among the targeted plants are crops (e.g., beet) and weeds (e.g., thistle and goosefoot). The images were acquired at different dates of May and June 2019, and different day times (see Figure 4b). Figures 6a and 7a show a RGB rendering of two of them with the D65 illuminant.
where |P j | is the number of pixels that characterize the considered patch. Among the 24 color patches of the ColorChecker chart, we use a learning subset P l of 12 patches for the learning procedure and the remaining 12 test patches P t for testing the quality of reflectance estimation (see Figure 5). The learning patches of P l are selected using an exhaustive search.
Among the 2,704,156 tested combinations, we retain the one that provides the lowest (mean absolute) reflectance estimation error (see Section 5.4).
The test subset P t is used to assess the performances that are reached by reflectance estimation methods, and the learning one P l is fed into a training-based reflectance estimation method, as described in the following.
Training-Based Reflectance Estimation
The linear Wiener (wn) estimation technique can be applied to estimate reflectance thanks to a learning procedure [23]. It is based on a matrix G that transforms radiance spectra into reflectance. From any radiance image I (B) in the database, we compute the spectrally-corrected vignetting-free radiance imageĨ (K) while using Equations (15) and (20), and then estimate the K-channel reflectance image as: To compute G, we use the spectra of the ColorChecker learning patches (P l subset) that are represented in each of our images. The estimation matrix G that is associated to each input radiance image is determined as: where T re f and T rad are the K × 12 matrices that are formed by horizontally stacking the centered and transposed reference reflectance vectors (fromR ) and radiance vectors (from the current imageĨ (K) ) of the learning patches, and denotes the transpose.
Evaluation Metrics
To evaluate the accuracy of reflectance estimation, we use the patches of the Col-orChecker test subset P t (see Figure 5). LetR (K) * ,P t j , * ∈ {rw, wa, ms, dwd, wn} denote the reflectance image that is estimated for patch P t j ∈ P t by either the proposed rw method (see Equation (28)) or the four implemented state-of-art methods (see Equations (23)-(30)). Black (24) Orange−yellow (12) Yellow (16) Red (16) Neutral8 (20) White (19) Cyan (18) Purple (10) Neutral65 (21) Blue − sky (3) Light − skin Bluish − green (6) Magenta (17) Blue − flower (5) Moderated − red (9) Green (14) Orange (7) Neutral35 (23) Yellow − green (11) Neutral5 (22) Foliage (4) Purple − Blue (8) Dark − skin (1) (b) Reflectance spectra of learning patches. (c) Reflectance spectra of test patches. This vector is compared to the reference reflectanceR [CC] of the same patch computed according to Equation (29). The spectra of the ColorChecker patches should be similar (and ideally superposed) to their laboratory counterparts when outdoor reflectance is well estimated. We objectively assess each estimated reflectance image thanks to the mean absolute error (MAE) and angular error ∆θ of each test patch P t j ∈ P t given by: and: where · 2 is the Euclidean norm. When ∆θ between two vectors (spectra in our case) is equal to zero, it means that these two vectors are collinear.
Results
We compute the mean absolute error MAE * and angular error ∆θ * averaged over all of the test patches of all reflectance images estimated from the whole database to obtain aggregated metrics. Table 1 presents the results for the five tested methods. The MAE and ∆θ are complementary metrics and they, respectively, highlight two important properties: the scale and shape of the estimated spectra. Indeed, while MAE is mainly sensitive to the scale of the estimated spectra, ∆θ especially focuses on the shape of the spectra, because it is a scale-insensitive measure. Consequently, there might be no correlation between the results that were obtained by the MAE measure and those obtained by ∆θ.
No method provides the best results according to the two metrics, as we can see from Table 1. Indeed, the wn and dwd methods provide better results than rw and wa according to the MAE, but the rw and wa methods provide better results in terms of ∆θ.
The ms method provides the worst results, because it only analyzes pixels of background and vegetation that strongly absorb the incident light in the visible domain. Hence, the biased illumination estimation in this domain affects the performance of ms method. It is worthwhile to mention that the wn method performance might also be biased, since it uses some of the ColorChecker patches as training references (to build estimation matrix G), while the other patches of the same chart are used to evaluate the reflectance estimation quality.
Among illumination-based methods that analyze a single reference device, rw provides similar results to wa in terms of ∆θ, as well as better MAE results. This shows that taking account of the illumination variation during the frame acquisitions improves the reflectance estimation quality.
Multispectral Image Segmentation
Now, we evaluate the contribution of our proposed rw-based reflectance estimation method for supervised crop/weed detection and identification. For this experiment, we focus on the beet (crop) that must be distinguished from thistle and goose-foot (weeds). First, vegetation pixels are detected and ground truth (labels) regarding vegetation pixels is provided by an expert in agronomy (Section 6.1). In order to evaluate the robustness of each considered feature against illumination conditions, we use a data set composed of 37 radiance (13 single-species and 24 mixed) images that we split into a learning and test set, denoted as S learn (23 images) and S test (14 images) (Section 6.2). The illumination conditions are various in the two sets and S test mostly includes images that are acquired on different days from those of S learn (see Figure 4). Note that, as a consequence, vegetation in the learning and test image sets may not be exactly at the same growth stages. We first compare the discrimination power of reflectance features provided by our rw method against radiance features to assess each reflectance estimation method for crop/weed identification and detection. Subsequently, we compare it with reflectance features that are estimated using each of the four considered state-of-the-art methods (wa, ms, wn, and dwd) (Sections 6.3-6.5).
Vegetation Pixel Extraction and Labelling
Only vegetation pixels are analyzed because we aim to detect/identify crops and weeds. They are distinguished from the background (white diffuser, ColorChecker, and soil pixels) using the normalized difference vegetation index (NDVI) [24]. We compute the NDVI values from the rw-based reflectance imageR (K) rw , since the rw method considers illumination variation, but the images provided by any other reflectance estimation method should yield similar vegetation pixel detection results. We consider p to be a vegetation pixel if its NVDI value is greater than a threshold γ: with the Snapscan "virtual" band centers λ 67 = 678.2 nm and λ 139 = 899.2 nm. Setting γ = 0.45 experimentally provides a good compromise between under-and oversegmentation of vegetation pixels. Noisy vegetation pixels are filtered out as much as possible by morphological opening. The vegetation pixels are then manually labelled by an expert in agronomy to build the segmentation ground truth for each multispectral image.
Learning and Test Vegetation Pixels
From the learning set S learn , we randomly extract N learning pixels per class. For a given class C i , i = [0, . . . , N C − 1], the number of extracted learning pixels per image depends on the number of images where class C i is represented in S learn (occurrences). Among the 23 learning images, the beet (crop) class appears in 17 images, thistle in nine images, and goosefoot in 12 images.
In the test set S test , beet, thistle, and goosefoot are represented, respectively, in 12, 10, and four images. For the weed detection task, we extract 2N learning pixels, half for crop and half for weed class. Because we merge thistle and goosefoot prototype pixels to build a single weed class, we extract N /2 learning pixels for thistle and N /2 for goosefoot. Each pixel is characterized by a K-dimensional (K = 141) feature vector of reflectance (or radiance) values. The reflectance/radiance images are averaged channel-wise over a 5 × 5 pixel window to reduce noise and within-class variability. Table 2 shows the number of learning and test pixels per class for weed detection and beet/thistle/goosefoot identification. All of the available pixels in S test are used to assess the generalization power of a supervised classifier.
Evaluation Metrics
The classical accuracy score can be a misleading measure to evaluate a classifier performance when the number of test pixels that are associated to each class is highly skewed (like in Table 2) [25] (p. 114). A classification model that predicts the majority class for all test pixels reaches a high classification accuracy. However, this model can also be considered as weak when misclassifying pixels of the minority classes is worse than missing pixels from the majority classes. In order to overcome this so-called "accuracy paradox", the performance of a classification model for imbalanced datasets should be summarized with appropriate metrics, such as precision/recall curve [25] (pp. 53-56, 114). Although some metrics may be more meaningful and easy to interpret, there is no consensus in the literature for choosing a single optimal metric. In our case, we want to correctly detect weed pixels without over-detection, because this would imply spraying crops with herbicides. Therefore, the performance of our classification model on both crop and weed detection/identification should be comparable. For this purpose, we use the per-class accuracy score and the weighted overall accuracy score. We also compute the F1-score that combines the precision and recall measures. These three measures should summarize the classification performance of imbalanced sets of test pixels well.
Let us denote the true test pixel labels as y and the set of predicted labels asŷ. The per-class accuracy score for class C i , i = [0, . . . , N C − 1], is: where y C ij andŷ C ij are the true and predicted labels for the j-th test pixel of class C i , respectively, and |y C i | is the number of test pixels of class C i . The weighted overall accuracy for binary and multiclass classifications is defined as: where ω C i = 1/|y C i | is the weight that is associated to class C i and computed as the inverse of its size, so as to handle imbalanced classes. Because the F1-score privileges the classification of true positives pixels (weed pixels in our case), we compute the overall F1-score as the population-weighted F1-score, so that the performances over all classes are considered: The F1-score of class C i is computed as: where and
Classification Results
The parametric LightGBM (LGBM) and non-parametric Quadratic Discriminant Analysis (QDA) classifiers are applied for supervised weed detection and identification problems. The choice of these two non-linear classifiers is motivated by their processing time during the learning and prediction procedures and their fundamentally different decision rules. Indeed, LGBM is a parametric tree-based classifier that requires a learning procedure to model a complex classification rule, whereas QDA is a simple non-parametric classifier that is based on Bayes' theorem to perform predictions. For LGBM, we retain the default parameter values (learning rate of 0.05, 150 leaves) and use the log loss function as the learning evaluation metric.
LGBM uses a histogram-based algorithm to bucket the features into discrete bins, which drastically reduces the memory and time consumption. The number of bins is set to 255 and the number of boosting operations to 100. Additionally, the feature fraction and bagging fraction parameters are set to 0.8 to increase LGBM speed and avoid over-fitting. Table 3 shows the classification results that were obtained with LGBM and QDA classifiers for each considered feature. Figures 6 and 7 show the color-coded vegetation pixel classification of two test images using the LGBM classifier in weed detection and identification tasks, respectively.
Let us first compare the classification performance of reflectance against radiance features for the weed detection task. From the results that are given in Table 3, we can see that reflectance features estimated by illumination-based methods (rw, wa, ms, and dwd) provide better classification results than radiance features in terms of the average F1 and accuracy scores, whatever the classifier. The worst classification results are obtained with reflectance features that are estimated using the wn method. Training-based methods, such as wn, can provide an accurate reflectance estimation of scene objects whose optical properties are close to those of the training samples. In our case, the optical properties of vegetation are very different from that of the training ColorChecker patches. Thus, wn provides inaccurate reflectance estimations at vegetation pixels, which affects its classification performance. Figure 6 illustrates the satisfying Accuracy and F1 scores that are obtained thanks to the analysis of illumination-based reflectance features by LGBM. Indeed, this figure shows that weed is globally well detected by these methods.
For weed identification, the classification performances of all the features are degraded, because they provide weak performances on the goosefoot class (see Figure 7). This lack of generalization might be due to the high within-class dispersion (since we consider vegetation at various growth stages) and/or the physiological vegetation changes. Let us now compare the classification performances of the reflectance features. The best overall classification results are obtained by our proposed rw method that performs well with both classifiers and reaches the highest average F1 and accuracy scores with LGBM for weed detection (85.4% and 86.1%, respectively). The wa method provides good classification results, better than those that were obtained by dwd, although the latter accounts for illumination variation during the frame acquisitions. The computation of illumination scaling factors to compensate for illumination variation may explain this poor performance, as well as the loss of spectral information (saturated reflectance values) in the near infra-red domain that is caused by illumination normalization (see Equation (A4)).
Experimental Conclusion
The experiments with this outdoor image database allow us to compare the performances of different reflectance features according to the estimation quality and pixel classification. The evaluation results are summarized by separately studying weed detection and identification. Tables 4 and 5 show the rank Rank , * obtained by each reflectance estimation method * according to each evaluation criterion used in Tables 1 and 3. The method with the lowest total rank is considered to be the best one, since it satisfies several criteria. The total ranks of ms and wn methods are the highest ones, because they provide the worst results for either estimation quality (ms) or classification performance (wn). On the one hand, the dwd method that uses two reference devices to cope with illumination variation provides the second best total rank for weed detection (see Table 4). On the other hand, the wa method that uses one reference device, but assumes that illumination is constant, gives the second best total rank for weed identification (see Table 5). Our rw method, which row-wise analyzes one single reference device in order to take account of illumination variation, reaches the best total ranks for both weed detection and identification problems. These experiments suggest that rw-based reflectance features are relevant for weed identification under variable illumination conditions. Their performance should also be confirmed with other crop species, such as bean and wheat.
Conclusions
This paper first proposes an original image formation model of linescan multispectral cameras, like the Snapscan. It shows how illumination variation during the multispectral image acquisition by this device impacts the measured radiance that is provided by a Lambertian surface element. Our model is versatile and it can be adapted to model the outdoor image acquisition of several multispectral cameras, such as the HySpex VNIR-1800 [26] or the V-EOS Bragg-grating camera used in [16]. From this model, we propose a reflectance estimation method that copes with illumination variation. Because such varying conditions may affect the reflectance estimation quality, we estimate illumination at the frame level using a row-wise (rw) approach. We experimentally show that the rw method is more robust against illumination variation than the state-of-the-art methods. We also show that rw-based features are more discriminant to target outdoor supervised weed detection and identification, and they provide the best classification results. The accuracy of weed recognition systems and their robustness against illumination can be improved using reflectance features. This allows for precision spraying techniques to be considered in order to get rid of weeds using fewer quantities of chemicals. This study enables to make a step towards sustainable agriculture. As future work, segmentation will be extended to other plant species (such as wheat and bean) and growth stages. Spectral feature selection and texture features extraction will also be studied to improve the crop/weed identification performance.
Each term in this equation is the average value over the row of p within the white diffuser subset W D in channel I b of either the full-field white diffuser image or the scene image. • Finally, the values ofR are normalized channel-wise to provide the dwd reflectance estimation as: where β b W P is the average value over the white patch subset W P in channelR b , and ρ b W D is the diffuse reflection factor of the white patch for the spectral band centered at λ b measured by a spectroradiometer in laboratory.
The B-channel reflectance imageR (B) dwd undergoes spectral correction and negative value removal (see Equations (15) and (16)) to provide the final K-channel reflectance imageR (K) dwd . | 10,806.6 | 2021-05-21T00:00:00.000 | [
"Mathematics"
] |
Role of DnaJ G/F-rich Domain in Conformational Recognition and Binding of Protein Substrates*
DnaJ from Escherichia coli is a Type I Hsp40 that functions as a cochaperone of DnaK (Hsp70), stimulating its ATPase activity and delivering protein substrates. How DnaJ binds protein substrates is still poorly understood. Here we have studied the role of DnaJ G/F-rich domain in binding of several substrates with different conformational properties (folded, partially (un)folded and unfolded). Using partial proteolysis we find that RepE, a folded substrate, contacts a wide DnaJ area that involves part of the G/F-rich region and Zn-binding domain. Deletion of G/F-rich region hampers binding of native RepE and reduced the affinity for partially (un)folded substrates. However, binding of completely unfolded substrates is independent on the G/F-rich region. These data indicate that DnaJ distinguishes the substrate conformation and is able to adapt the use of the G/F-rich region to form stable substrate complexes.
The Hsp40 protein family, also called JDP family, exhibits a wide structural and functional diversity (1,2). All members of this family share a ϳ70-amino acid J-domain that allows them to cooperate with Hsp70 chaperones in different processes essential for cell life. The Hsp40 family can be divided into three groups depending on domain composition (3). Type I Hsp40s are constituted by a J-domain followed by a G/F-rich flexible linker, a zinc-binding domain and a conserved C-terminal domain involved in substrate binding and dimerization (4,5). Members of this group are well-known proteins such as Escherichia coli DnaJ (6) and Saccharomyces cerevisiae Ydj1p (7). Type II proteins, e.g. yeast Sis1p (8), are similar to type I but lack the zinc finger domain. Although structurally and functionally related, these two groups contain unique protein modules (8,9) and have different quaternary structures (10). Type III JDPs contain only the conserved J-domain that can be located anywhere in the protein sequence. To this group belong functionally diverse proteins such as E. coli DjlA, the clathrin-uncoating auxilin (11), mitochondrial Tim14 and Tim16 (12), and endoplasmic reticulum Sec63p (13).
The functionality of Hsp40 proteins is based on their ability to regulate the ATPase activity of Hsp70 chaperones, and to target protein substrates to their Hsp70 partners. By enhancing the ATP hydrolysis rate of Hsp70, Hsp40s facilitate the conformational transition to the high substrate affinity state of the chaperone ensuring tight binding of the polypeptide chain. The activation of E. coli DnaK intrinsic ATPase by DnaJ and the synergistic effect observed in the presence of substrates are strictly dependent on the J-domain and the G/F-rich region (14,15). Mutations of a conserved HPD motif in the J-domain abolish the functional interaction with Hsp70 (14,16,17). The second essential function of Hsp40s is the binding and delivery of substrate proteins to Hsp70 (18,19), a process not well understood. Type I Hsp40s can act as independent chaperones and bind and suppress aggregation of non-native polypeptides (20,21), while type II proteins have to cooperate with Hsp70s to prevent aggregation (22). The crystal structure of Ydj1p and Sis1p reveals a hydrophobic pocket in the C-terminal domain that can accommodate short peptides (23,24). DnaJ has a preference for peptides enriched in aromatic and large hydrophobic residues (25). Binding of substrates by type I Hsp40s also depends on the Zn-binding domain (20,21). The binding/delivery of substrates to Hsp70s and the stimulation of its ATPase activity are not independent processes and work in a synergic manner.
The aim of this study is to gain knowledge on the interaction of E. coli DnaJ with protein substrates in different conformational states (folded, partially folded, and unfolded). We found that DnaJ distinguishes completely unfolded from partially folded or folded conformations. Of particular importance to recognize these conformations is the G/F-rich domain, which is required for the cochaperone to display a physiologically significant affinity for a folded substrate as RepE, and for folding intermediates of the thermal unfolding pathway of malate dehydrogenase (MDH) 4 and luciferase, but not for chemically unfolded luciferase.
EXPERIMENTAL PROCEDURES
Protein Cloning, Expression, and Purification-DnaJ, DnaJ ⌬73 , and DnaJ ⌬107 were amplified by standard PCR techniques and cloned into pET22 vector. Wt DnaJ and mutants were expressed in BL21(DE3) cells and purified as described (26). It should be noted that following the mentioned protocol a small amount of endogenous wt DnaJ could copurify with the mutants. However it was not detected in Coomassie-stained SDS-PAGE (supplemental Fig. S1A). Additionally, the ATPase activity of DnaK was not stimulated by any of the truncation mutants in contrast to wt DnaJ (supplemental Fig. S1B). Thus, the activity (or lack of) of DnaJ ⌬73 and DnaJ ⌬107 was not affected by this contamination. DnaK, GrpE, and ClpB were obtained as described (27)(28)(29). f-RCMLA was produced from ␣-lactalbumin (Sigma) (30) and fluorescently labeled using 5Ј-carboxyfluorescein succinimidyl ester (Invitrogen Molecular Probes) (31). MDH and luciferase were purchased from Sigma and Roche, respectively. Anti-DnaJ antibody was purchased from StressMarq.
RepE expression vector was a kind gift from Prof. D. Bastia, and the protein was purified according to the published protocol (32). For pulldown experiments, an N-terminal His-tagged version of RepE was used. RepE was cloned into pHAT vector and transformed in BL21(DE3) cells. Cell cultures in exponential phase were cooled down to 20°C and induced with 0.1 mM isopropyl-1-thio--D-galactopyranoside. Cells were harvested after 24 h of incubation at 20°C, resuspended in 40 mM sodium phosphate, pH 8.0, 300 mM NaCl, 10 mM imidazole, 1 mM PMSF, 10% glycerol, and lysed by sonication. After centrifugation at 35,000 rpm for 40 min at 4°C, 2 ml NiNTA beads (Quiagen) were added to the supernatant, and the protein was allowed to bind for 2 h at 4°C. The NiNTA beads were washed with a high salt buffer and the protein was eluted with a 20-ml linear gradient to 500 mM imidazole. RepE-containing fractions were pooled, diluted 5-fold with 40 mM sodium phosphate, pH 6.8, 10% glycerol and loaded in a HiTrap SP 5 ml column (GE Healthcare) equilibrated in 40 mM sodium phosphate, pH 6.8, 100 mM NaCl, 0.1 mM EDTA, 2 mM -mercaptoethanol, 10% glycerol. The column was eluted with a 100 ml linear gradient, with equilibration buffer containing 1 M NaCl. HisRepE was concentrated and stored at Ϫ70°C.
Anisotropy Measurements-RepE (1 M) and f-RCMLA (1 M) were incubated with increasing concentrations of wt DnaJ or the deletion mutants (0.1-40 M) in 20 mM Hepes/KOH, pH 7.4, 50 mM KCl, 5 mM DTT, 0.1 mM EDTA. Samples were equilibrated overnight at 4°C following incubation for 1 h at 25°C before measurements. Fluorescence anisotropy measurements were performed in a Fluorolog spectrofluorimeter (Jobin Ibon) with excitation and emission wavelengths at 280 nm and 345 nm for RepE, and 494 and 518 nm for f-RCMLA. Slit widths were 8 nm. The fraction of RepE bound to DnaJ was calculated and the experimental points were fitted by a non-linear least squares method to a single quadratic equation assuming one binding site.
Trypsin Partial Proteolysis-Proteolysis experiments were performed at 30°C in 25 mM Hepes/KOH, pH 7.6, 50 mM KCl, 10 mM -mercaptoethanol. 20 M DnaJ, and 20 M RepE were incubated for 2 h at 25°C to form the complex. Proteolysis was initiated by addition of trypsin at 1:200 or 1:40 (w/w) ratios. Reactions were allowed to occur for different times at 30°C and stopped by addition of 1 mM PMSF. As a control, the isolated proteins were also treated with trypsin at the same ratios. Proteolysis products were analyzed by SDS-PAGE (12.5% gels) followed by staining with Coomassie Brilliant Blue.
Partial tryptic digestions were analyzed by LC-MS in a Q-Tof Micro (Waters) mass spectrometer interfaced with a CapLC System (Waters). Tryptic digestion was stopped by the addition of formic acid at a final concentration of 0.5% and loaded onto a Symmetry 300 C4 NanoEase Trap precolumn (Waters). The precolumn was connected to a Symmetry 300 C4 analytical column (75 m ϫ 150 mm, 3.5 m, Waters) equilibrated in 5% acetonitrile and 0.1% formic acid. Polypeptides were eluted with a 30 min linear gradient of 5-60% acetonitrile directly onto a NanoEase Emitter (Waters) and mass spectra acquisition was performed in MS mode. Mass spectra were manually inspected and combined, and intact mass of the polypeptides was determined by MaxEnt1 software (Waters) using default deconvolution parameters. Mass ranges were selected based on protein sequence information and software was set to iterate to convergence. Experimentally obtained masses were matched to DnaJ and RepE protein sequence fragments using the BioLynx tool embedded in MassLynx 4.1 software (Waters).
Pulldown Experiments-For pulldown of wt DnaJ and mutants, a N-terminal His-tagged version of RepE was used. HisRepE (10 M) was incubated with NiNTA beads for 90 min at 4°C with gentle shaking in 20 mM Hepes/KOH, pH 7.4, 50 mM KCl, 0.5 mM MgCl 2 , 50 mM imidazole. DnaJ or mutants (10 M) were added and incubated for 90 min at 25°C. Unbound proteins were removed by washing three times, and pellets were analyzed by SDS-PAGE.
MDH and Luciferase Aggregation and Refolding-Thermal unfolding of 1 M MDH and 0.25 M luciferase was induced by incubation at 47°C and 42°C, respectively, in 20 mM Hepes/ KOH, pH 7.4, 50 mM KCl, 5 mM MgCl 2 , 5 mM DTT in the absence or the presence of DnaJ. Unfolding of luciferase (25 M) was achieved by incubation in 6 M GndHCl, 30 mM Tris/ HCl, pH 7.5, 5 mM DTT for 2 h at 22°C. The unfolded protein was diluted 100-fold in the above buffer in the absence or presence of DnaJ or DnaJ mutants (0.1-5 M). Aggregation was monitored by continuous measurement of light scattering at 500 nm (MDH) or 320 nm (luciferase) in a Fluorolog spectrofluorimeter (Jobin Ibon).
Refolding experiments were performed after denaturation of MDH (1 M) for 20 min at 47°C or luciferase (80 nM) for 30 min at 42°C, in the absence or presence of 1 M DnaJ or DnaJ mutants in the above buffer. For chemical denatured substrates, luciferase (2.5 M) was incubated in 6 M GndHCl, 30 mM Tris/ HCl, pH 7.5, 5 mM DTT for 2 h at 22°C, and diluted 100-fold in the above buffer in the absence or presence of 1 M DnaJ or DnaJ mutants. Reactivation was achieved by addition of 2 mM ATP, an ATP-regeneration system (20 ng ml Ϫ1 pyruvate kinase, 4 mM PEP), and different combinations of 1 M DnaK, 1 M GrpE, and 1.5 M ClpB. Samples were incubated for 2 h at 30°C, and MDH and luciferase activity were measured as described (33,34).
Intrinsic Fluorescence-Luciferase intrinsic fluorescence spectra were recorded in a Fluorolog spectrofluorimeter (Jobin Ibon) using 295 nm as excitation wavelength and 8 nm slit widths. Luciferase concentration was 0.25 M, and the buffer was 20 mM Hepes/KOH, pH 7.4, 50 mM KCl, 5 mM MgCl 2 , 5 mM DTT. For thermal denaturation, the sample was heated to 42°C for 20 min in the absence or presence of 5 M DnaJ. For chemical denaturation, 25 M luciferase was denatured in 6 M GndHCl, 30 mM Tris/HCl, pH 7.5, 5 mM DTT for 2 h at 22°C, and diluted 100-fold in the above buffer containing 6 M GndHCl or 0.5 M DnaJ.
F-APPY Binding Assays-F-APPY (50 nM) was incubated overnight at 4°C in the absence of chaperones or with DnaK (1 M), DnaJ (1 M), or DnaJ ⌬107 (1 M). As a control, recombinant nucleoplasmin (1 M) was used. Buffer was 20 mM Hepes/ KOH, pH 7.4, 50 mM KCl, 5 mM MgCl 2 . Fluorescence was measured using 492 nm and 516 nm as excitation and emission wavelengths, respectively. The fluorescence intensity of a freshly dissolved peptide solution was set as 100%. For separation in density gradients, F-APPY (0.2 M) was incubated with 10 M DnaJ or DnaJ ⌬107 for 1 h at 25°C and loaded on top of a 5-30% sucrose gradient. After 22 h centrifugation at 12,600 ϫ g at 4°C, the gradient was fractionated using a PGFip (Biocomp). F-APPY was detected measuring the fluorescence intensity of the samples, and DnaJ and DnaJ ⌬107 by SDS-PAGE.
DnaJ Binds RepE with High
Affinity-First, we tested the ability of DnaJ to bind soluble protein substrates such as RepE or reduced carboxymethylated lactalbumin (RCMLA). RepE is the initiation factor of plasmid F and is functionally homologous to RepA of P1 plasmid, which was shown to be a high affinity DnaJ substrate (35). Depending on its oligomeric state, RepE acts as dimer/repressor or monomer/activator of plasmid replication (36). The conversion from dimers to monomers depends on the action of DnaK/DnaJ/GrpE (36). RCMLA adopts a partially folded soluble conformation that makes it a suitable substrate for chaperones (6). Complex formation was first followed by the increase in fluorescence anisotropy of the substrates. RepE-DnaJ interaction was monitored by the intrinsic fluorescence of the substrate, taking advantage of the lack of tryptophan residues in the DnaJ sequence. In the case of RCMLA, the substrate was labeled with fluorescein. As shown in Fig. 1A, DnaJ had a different affinity for the two substrates: while it bound RepE with a K d ϭ 5.2 Ϯ 0.7 M, the affinity for RCMLA was much lower, in accordance with previous observations (6,35). The stability of the DnaJ-RepE complex was also demonstrated by chemical cross-linking with glutaraldehyde ( Fig. 1B). Other crosslinkers as BS3 and DSP were also tried, however best results were obtained with glutaraldehyde. Cross-linking of isolated DnaJ or RepE resulted in adducts corresponding to the dimeric forms of both proteins. When preformed DnaJ:RepE complexes were subjected to cross-linking, one adduct, that could be recognized by a DnaJ antibody (Fig. 1B, lower panel), with apparent molecular mass compatible with a cross-link between DnaJ and RepE was also observed. Furthermore, the complex between DnaJ and RepE could be detected by pulldown experiments with NiNTA using an N-terminal Histagged version of RepE (see below).
RepE Protects Specific Tryptic Sites in DnaJ-We used trypsin partial proteolysis to investigate DnaJ sites that were interacting with RepE. This approach assumes that RepE binding would prevent access of trypsin to specific sites within the DnaJ molecule. Isolated DnaJ and RepE, and the complex formed by both proteins were treated with trypsin at low (1:200) and high (1:40) molar ratios for different times. Proteolytic fragments were resolved by SDS-PAGE. To identify the fragments, samples of isolated DnaJ and RepE were digested with trypsin at 1:40 molar ratio and analyzed by mass spectrometry (MS). The presence of DnaJ did not significantly affect RepE proteolytic pattern at any trypsin concentration, however the tryptic pattern of DnaJ was modified upon complex formation with RepE. At low trypsin molar ratios ( Fig. 2A), DnaJ was rapidly degraded in the absence of RepE mainly into fragments of apparent molecular mass of ϳ9 and 28 -29 kDa, which can be assigned to the N-terminal J domain, and C-terminal fragments cleaved at positions Arg-108, Arg-110, and Arg-113, just after the G/F-rich region OCTOBER 29, 2010 • VOLUME 285 • NUMBER 44 (Table 1 and Fig. 2C). In the presence of RepE, full-length DnaJ (marked with an arrow) is protected against proteolysis, indicating that cleavage sites located after the G/F-rich region were not accessible to the protease ( Fig. 2A). At 1:40 trypsin molar ratios (Fig. 2B), the 28 -29 kDa C-terminal fragments were further degraded in the absence of RepE, giving rise to fragments of 20 -22 kDa originated from cleavage within the Zn finger region (Domain II) at positions Lys-153, Arg-173, and Arg-189 (Fig. 2C). These tryptic sites were also protected when DnaJ was complexed to RepE, and degradation of 28 -29 kDa fragments was severely impaired (Fig. 2B). In summary, these results suggest that RepE very likely contacts the G/F-rich region and the Zn-binding domain II of DnaJ (Fig. 2C); thus these domains might play an important role in DnaJ-substrate complex formation. Note the absence of tryptic sites within the G/F-rich region that makes impossible the formation of proteolytic fragments starting at this domain.
DnaJ-Substrate Interactions
The G/F-rich Region Is Required for Stable Binding of Substrates-To investigate the role of the N-terminal domains of DnaJ on substrate binding, we produced two deletion mutants, DnaJ ⌬73 and DnaJ ⌬107 , which lack the J-domain, and J-domain and G/F-rich region, respectively. As previously found (37), none of these deletions compromised the overall stability of the protein as seen by circular dichroism, which showed secondary structure and denaturation temperatures similar to wt DnaJ (supplemental Fig. S2, A and B). The interaction of both mutants with RepE was investigated by fluorescence anisotropy and pull-down experiments. As shown in Fig. 3A, DnaJ ⌬73 bound RepE with similar affinity to wt DnaJ, while RepE anisotropy did not significantly increase in the presence of DnaJ lacking the J-domain and G/F-rich region within the same concentration range. It should be mentioned that DnaJ ⌬73 had a tendency to aggregate at high concentrations, possibly because of the exposure of the G/F-rich domain to the solvent. As a consequence, higher anisotropy values were observed with this mutant at concentrations above 10 M. These data suggest that the G/F-rich domain is required to form a high affinity complex with RepE. Pull-down experiments with His-tagged RepE confirmed these results (Fig. 3B), because DnaJ ⌬107 , in contrast to wt DnaJ and DnaJ ⌬73 , was unable to coprecipitate with RepE above background levels, even when protein concentration was raised to compensate the effect of its reduced affinity.
The role of the G/F-rich region in the stable binding of protein substrates was also investigated using unfolded MDH and luciferase. First, aggregation of thermally denatured MDH and luciferase was followed by light scattering in the absence or presence of DnaJ or of the deletion mutants (Fig. 4, A and B). In both cases DnaJ hindered aggregation in a concentration-dependent manner, higher DnaJ:substrate molar ratios being required to protect luciferase, i.e. equimolar DnaJ significantly reduced aggregation of MDH, and this was completely abolished at 5:1 molar ratio, while a 20:1 DnaJ molar excess was required to completely protect luciferase. These results might reflect a different affinity of DnaJ for these unfolded substrates as previously observed with RepE and RCMLA (this work; Ref. 35). The ability of DnaJ ⌬73 to protect thermally denatured MDH and luciferase against aggregation was similar to that of wt DnaJ (not shown), in contrast to that of DnaJ ⌬107 , which was severely impaired. DnaJ ⌬107 slightly protected MDH against aggregation at high molar ratios (10:1) and slowed down the kinetics of luciferase aggregation, suggesting that this deletion mutant may interact with these partially (un)folded substrates with a markedly reduced affinity, in agreement with results previously found with Ydj1p (21). Interestingly, a different scenario was found when luciferase was denatured with guanidinium hydrochloride (GndHCl) and substrate aggregation was induced by dilution in the absence or presence of chaperones. Under these conditions wt DnaJ and DnaJ ⌬107 hindered similarly the aggregation of unfolded luciferase (Fig. 4C). These results suggest that the requirement of the G/F-rich region to bind unfolded substrates is complex and depends on the conformational properties of the substrate protein. Whereas GndHCl extensively denatured the protein ( em ϭ 364 nm), thermal denaturation at 42°C induced a partially (un)folded state ( em ϭ 352 nm). The fluorescence properties of the DnaJluciferase complexes were similar ( em ϭ 353 nm), regardless of the denaturation method used, and different from that of native luciferase ( em ϭ 345 nm). Therefore, DnaJ would interact with refolding intermediates of GndHCl-denatured luciferase that would show a larger exposure of tryptophan residues to the solvent ( em varies from 364 to 352 nm), than those that populate its thermal unfolding pathway ( em varies from 345 to 352 nm), and thus that will be more extensively unfolded. The ability of DnaJ and the deletion mutants to protect aggregation of thermally denatured MDH and luciferase was also tested by assaying substrate refolding. After denaturation of the substrates, as indicated, in the absence or the presence of different DnaJ variants, refolding was started by addition of different combinations of DnaK/GrpE/ClpB and ATP, and incubation of the sample for 2 h at 30°C (Fig. 4, D and E). Refolding of MDH and luciferase denatured in the absence of chaperones was significant only when the DnaK system and ClpB cooperated, while when 1 M DnaJ was used to protect aggregation, the DnaK system reactivated MDH and luciferase on its own, albeit to a lower extent. A higher refolding yield was obtained when ClpB was also used in the refolding step in the presence of equimolar DnaJ, reflecting the lower aggregation of the substrate protein, as suggested by light scattering data (Fig. 4, A and B). When the substrates were denatured in the presence of DnaJ ⌬73 , the lack of the J-domain hampered reactivation by DnaK/GrpE or DnaK/GrpE/ClpB, despite the protection effect exerted by this mutant as described above. Addition of wt DnaJ together with DnaK/GrpE or DnaK/GrpE/ClpB could reactivate MDH and luciferase denatured in the presence of DnaJ ⌬73 with lower yields, suggesting that the mutant competed with wt DnaJ for substrate binding and, therefore, lowered the final reactivation yield. In contrast, when MDH and luciferase were aggregated in the presence of DnaJ ⌬107 , refolding could only be observed when a combination of DnaK/DnaJ/ GrpE/ClpB was used, the yield being similar to that of the control experiment. This confirms that under these conditions DnaJ ⌬107 does not interact with or protect aggregation of partially (un)folded MDH and luciferase.
Finally, refolding of chemically denatured luciferase was performed. Aggregates of chemically denatured luciferase were easily reactivated by K/J/E and addition of ClpB did not improve significantly the refolding yield. When DnaJ was present in the dilution buffer, aggregation of luciferase was prevented (Fig. 4C) and higher refolding yields were obtained. DnaJ ⌬73 or DnaJ ⌬107 also avoided aggregation, when present in the dilution step (not shown and Fig. 4C). Because the J-domain is missing in both mutants, refolding could only be achieved when wt DnaJ was added, however the yields were significantly lower (around 20%), meaning that both mutants were competing with wt DnaJ for the substrate as observed above. These results support that DnaJ ⌬107 was able to bind chemically denatured luciferase and prevent its aggregation. OCTOBER 29, 2010 • VOLUME 285 • NUMBER 44
Binding of Peptides Is Not Affected by the N-terminal Deletion-
Although DnaJ ⌬107 was correctly folded, as mentioned above, it was possible that subtle changes in its tertiary structure affected the stability of the peptide binding site. To investigate this possibility, DnaJ ⌬107 was treated with trypsin (1:200 for 30 s), and the fragments obtained analyzed by MS (supplemental Fig. S2C). We found that the deletion did not increase the accessibility to trypsin, and digestion of DnaJ ⌬107 gives rise to fragments starting at positions 110 and 113, similar to those found for wt DnaJ. Therefore the stability of the adjacent -strand starting at Leu-117 (Ile-116 in Ydj1p), which forms part of the binding pocket, was not affected by the N-terminal deletion. Next, we tested binding of F-APPY peptide (fluorescein-CALLQSRLLLSAPRRAAATARY) to DnaJ and DnaJ ⌬107 . F-APPY binds Hsp70 proteins with high affinity (34,38) and contains a sequence similar to peptide p5Ј, known to be a good DnaJ binder (39). Unfortunately neither a significant change in peptide fluorescence nor anisotropy enhancement was observed after binding to DnaJ and, therefore, the K d of the association reaction could not be estimated. The hydrophobicity of F-APPY promotes its aggregation and only a small percentage of the fluorescence of a freshly dissolved peptide was observed after incubation for 12 h at 4°C (Fig. 5A). However, in the presence of DnaK, DnaJ, and DnaJ ⌬107 , but not of an unrelated protein as nucleoplasmin, aggregation was abolished indicating that F-APPY was bound to the chaperones and thus remained in solution. When isolated F-APPY was loaded in a sucrose density gradient, the peptide was found in two fractions: one, more abundant, at the bottom and a second one at the top (fraction 1) of the gradient corresponding to aggregated and soluble peptide populations, respectively (Fig. 5B). In the presence of chaperones, aggregation was hindered, as mentioned above, and F-APPY co-sedimented around fraction 5 with DnaJ and DnaJ ⌬107 . These results demonstrate that the N-terminal deletion does not affect the integrity of DnaJ peptide binding site.
DISCUSSION
Chaperones of the Hsp70 and Hsp40 families assist many essential cellular processes due to their ability to interact with (un)folded proteins and remodel their conformation. Hsp40 proteins act as cochaperones of Hsp70s, modulating their function by two mechanisms: (i) activation of the intrinsically weak ATPase activity of the chaperone that depends on the conserved J domain, present in all Hsp40 proteins; and (ii) delivery of protein substrates, a poorly understood process. In some cases, Hsp40 proteins can act independently and form stable complexes with folded and unfolded polypeptides, avoiding their aggregation. However the way they interact with substrates of different conformational properties remains largely unexplored. Type I and Type II Hsp40s, as E. coli DnaJ and S. cerevisiae Ydj1p and Sis1p, contain a C-terminal domain able to interact with unfolded polypeptides (23,24). The role of this domain in substrate binding is critical and mutations of residues that form the binding pocket impair the correct function of the protein (4,22,40). Hsp40s are modular proteins, and in addition to the C-terminal region, other domains have been involved in the interaction with some substrates (20,21,41). Here we have investigated the interaction of E. coli DnaJ with folded and (partially) unfolded substrates, focusing on the role of the N-terminal J and G/F-rich domains. We find that DnaJ has a marked substrate specificity as it forms a high-affinity complex with RepE while it interacts with low affinity with RCMLA, in agreement with previous findings (35). DnaJ has a preference to bind peptide stretches of ϳ8 residues enriched in aromatic and large aliphatic residues (25). Such sequences are found in both RepE and RCMLA, thus it remains unclear whether the higher affinity for RepE is due to the greater exposure of a sequence motif or to a specific folding pattern.
Our data indicate that RepE could directly contact the G/Frich region and Zn-binding domain of DnaJ. The involvement of the Zn-binding domain in the binding of several substrates has been put forward for DnaJ and other Type I Hsp40s as the highly homologous yeast Ydj1p and endoplasmic reticulum ERdj3 (21,41,42). Here we give evidence that a substrate as RepE might also interact with this DnaJ domain, as observed by protection against proteolysis of several tryptic sites located within this domain. On the other hand, the role of the G/F-rich region in substrate binding seems controversial. Several reports have shown that this domain is dispensable since deletion or mutation of DIF motifs within this DnaJ domain did not modify the ability to bind either 32 or chemically unfolded luciferase (15,43). Similarly, Type I ERdj3 and Type II Sis1p bound several substrates independently of this domain (42,44). Our data suggest that the requirement of the G/F-rich domain might depend on substrate conformation. Thus DnaJ ⌬107 interacts with chemically unfolded luciferase with similar affinity than wt DnaJ, resulting in very similar yields of protection against aggregation. However, this deletion mutant cannot efficiently protect aggregation of thermally unfolded luciferase and MDH within the concentration range used, suggesting a decreased affinity for these partially (un)folded conformations. The lower aggregation rate observed at low mutant concentrations might indicate a transient non-stable interaction with the substrate that cannot avoid the aggregation process. These data would suggest that DnaJ differentiates between chemically and thermally denatured luciferase and adapts the use of the G/F-rich domain to stably bind the latter. The main difference between the two denatured conformations of luciferase is the degree of unfolding of the polypeptide chain. Because GndHCl extensively denatures the protein, the cochaperone will probably bind more unordered conformations than during thermal denaturation of the substrate, where luciferase will gradually unfold in the presence of DnaJ, as observed by intrinsic fluorescence. Furthermore, DnaJ ⌬107 also fails to protect aggregation of thermally denatured MDH. As an indirect method to estimate the apparent affinity of wt DnaJ and DnaJ ⌬107 for thermally denatured MDH, we have used the percentage of protection against aggregation (not shown). Whereas the estimated K d for wt DnaJ is 0.45 M, the affinity drops at least one order of magnitude for DnaJ ⌬107 , making impossible the precise estimation of its value within the protein concentration range used. In agreement with our results, deletion of the G/F-rich domain of Ydj1p also resulted in a lower efficiency to suppress aggregation of rhodanase (21). Finally, the G/F-rich domain is strictly required to bind a folded substrate as RepE, the affinity being drastically reduced, by above two orders of magnitude, for the deletion mutant. Taken together, these results suggest that DnaJ might require a larger interaction surface to bind a folded substrate as RepE, which would be provided by the G/F-rich region and possibly by the Zn-binding domain. Deletion of this domain has been shown to similarly reduce the affinity for both folded (such as the homologous protein RepA and P) (41) and unfolded (chemically denatured rhodanase) (20,21) substrates. Thus, the ability to differentiate the conformation of the substrate protein could mainly depend on the G/F-rich region. A recent study has shown that DnaJ interacts with side chains of a protein substrate in contrast to DnaK, its Hsp70 partner, that establishes contacts with the backbone (45). Binding of unfolded substrates would allow the interaction of DnaJ peptide binding site with exposed hydrophobic side chains, as seen for short apolar peptides, thus making the complex stable in a polar environment. However, as the conformation of the interacting substrate is more stable, or "native," the hydrophobic side chains would be hidden in the apolar core of the protein, thus the interactions to stabilize the DnaJ-substrate complex would become more polar. As compared with hydrophobic interactions, the number of polar contacts required to stabilize FIGURE 5. DnaJ ⌬107 is able to bind short hydrophobic peptides. A, aggregation of F-APPY (50 nM) after 12 h of incubation at 4°C is prevented by DnaK, DnaJ, and DnaJ ⌬107 , but not by nucleoplasmin, a histone chaperone. The fluorescence of a freshly dissolved peptide was set as 100%. Chaperone concentration was 1 M. B, F-APPY (200 nM) alone or in the presence of DnaJ or DnaJ ⌬107 (10 M) was loaded on a sucrose gradient. After 22 h centrifugation, the gradient was fractionated, and F-APPY was detected by fluorescein fluorescence intensity (upper panel) and proteins by SDS-PAGE analysis (lower panel) of the fractions. the complex would be most likely larger, increasing in parallel with the interacting surface that provides them. This interpretation would explain why the difference in apparent affinities between wt DnaJ and DnaJ ⌬107 is negligible for chemical unfolded luciferase, around an order of magnitude lower for intermediates of MDH and luciferase thermal unfolding pathways, and two orders of magnitude for a folded substrate as RepE. However, binding of native 32 to DnaJ seems to contradict this interpretation, since 32 -DnaJ complex does not rely on the presence of the G/F-rich region and Zn-binding domain (15,41). This apparent disagreement could be rationalized considering that 32 adopts a loosely folded and highly flexible conformation, as amide hydrogen exchange has demonstrated (46).
Deletion of DnaJ and Sis1p G/F-rich domain in E. coli and yeast cells, respectively, exhibits a poisonous effect (43,44). This effect was attributed to a possible role of the G/F-rich domain in the "targeting" of substrates to DnaK/Ssa1p and failure to activate the chaperone, which might result in kinetically trapped chaperone-substrate complexes (43,44). As suggested (43), some of these substrates might be essential proteins for cell survival. In this context, we show that direct binding of several folded and partially folded substrates strictly depends on the G/F-rich region. For substrates that require the G/F-rich domain to bind to DnaJ, this protein domain most likely participates directly in the transfer of substrates to DnaK and activation of the chaperone ATPase activity, possibly by correct positioning of the adjacent J-domain. For other substrates that bind to DnaJ independently of this protein region, the G/F-rich domain might also participate in complex stabilization, through transient interactions with the polypeptide chain.
In summary, the data presented here indicate that DnaJ combines different domains to interact with protein substrates depending on their conformational properties. Whereas binding of small apolar peptides or unfolded proteins does not rely on the presence of the G/F-rich region, complex formation with partially (un)folded or folded substrates requires this protein domain. Therefore the cochaperone can, if necessary, provide an interacting surface to stably bind specific substrates, besides the "binding site" adapted to interact with apolar amino acid stretches at the C-terminal domain. The generalization of this interpretation to rationalize the interaction of protein substrates with DnaJ would require a systematic analysis of more substrates. However the number of native substrates available is rather limited, and only two proteins (RepA, a RepE homolog, and 32 ) have been described as high-affinity DnaJ binders. This study points to the importance of the substrate conformational properties in the interaction with DnaJ, which in turn determines the cochaperone domains involved in the formation of stable complexes. | 7,732.2 | 2010-08-20T00:00:00.000 | [
"Biology",
"Chemistry",
"Physics"
] |
A Nonlocal Cauchy Problem for Fractional Integrodifferential Equations
This paper is concerned with a nonlocal Cauchy problem for fractional integrodifferential equations in a separable Banach space X. We establish an existence theorem for mild solutions to the nonlocal Cauchy problem, by virtue of measure of noncompactness and the fixed point theorem for condensing maps. As an application, the existence of the mild solution to a nonlocal Cauchy problem for a concrete integrodifferential equation is obtained.
Introduction
Nonlocal Cauchy problem for equations is an initial problem for the corresponding equations with nonlocal initial data.Such a Cauchy problem has better effects than the normal Cauchy problem with the classical initial data when we deal with many concrete problem coming from engineering and physics cf., e.g., 1-10 and references therein .Therefore, the study of this type of Cauchy problem is important and significant.Actually, as we have seen from the just mentioned literature, there have been many significant developments in this field.
On the other hand, fractional differential and integrodifferential equations arise from various real processes and phenomena appeared in physics, chemical technology, materials, earthquake analysis, robots, electric fractal network, statistical mechanics biotechnology, medicine, and economics.They have in recent years been an object of investigations with much increasing interest.For more information on this subject see for instance 9, 11-18 and references therein.c D q x t Ax t f t, x t t 0 k t, s h t, s, x s ds, t ∈ 0, T , where k and g are given functions to be specified later and the fractional derivative is understood in the Caputo sense, this means that, the fractional derivative is understood in the following sense: c D q x t : L D q x t − x 0 , t > 0, 0 < q < 1, 1.3 and where L D q x t : 1 Γ 1 − q d dt t 0 t − s −q x s ds, t > 0, 0 < q < 1 1.4 is the Riemann-Liouville derivative of order q of x t , where Γ • is the Gamma function.
Our main purpose is to establish an existence theorem for the mild solutions to the nonlocal Cauchy problem based on a special measure of noncompactness under weak assumptions on the nonlinearity f and the semigroup {T t } t≥0 generated by A.
Existence Result and Proof
As usual, we abbreviate u L p 0,T , R with u L p , for any u ∈ L p 0, T , R .
As in 16, 17 , we define the fractional integral of order q with the lower limit zero for a function f ∈ AC 0, ∞ as provided the right side is point-wise defined on 0, ∞ .Now we recall some very basic concepts in the theory of measures of noncompactness and condensing maps see, e.g., 19, 20 .
Definition 2.1.Let E be a Banach space, 2 E the family of all nonempty subsets of E, A, ≥ a partially ordered set, and α : 2 then we say that α is a measure of noncompactness in E.
Definition 2.2.Let E be a Banach space, and If for every bounded set Ω ⊆ Y which is not relatively compact, then we say that F is condensing with respect to the measure of noncompactness α or α-condensing .
be a one-sided stable probability density, and For any z ∈ X, we define operators {Y t } t≥0 and {Z t } t≥0 by Y t z ∞ 0 ξ q σ T t q σ zdσ, Z t z q ∞ 0 σt q−1 ξ q σ T t q σ zdσ.
2.8
If 2.16 Proof.First of all, let us prove our definition of the mild solution to problem 1.2 is well defined and reasonable.Actually, the proof is basic.We present it here for the completeness of the proof as well as the convenience of reading.Write
2.17
Clearly, the nonlocal Cauchy problem 1.2 can be written as the following equivalent integral equation: provided that the integral in 2.18 exists.Formally taking the Laplace transform to 2.18 , we have Therefore, if the related integrals exist, then we obtain
2.20
Now using the uniqueness of the Laplace transform cf., e.g., 21, Theorem 1.1.6, we deduce that ξ q σ T t − s q σ a x s dσ ds.
2.21
Consequently, we see that the mild solution to problem 1.2 given by Definition 2.3 is well defined.
Next, we define the operator F : C 0, T , X → C 0, T , X as follows: It is clear that the operator F is well defined.
The operator F can be written in the form F F 1 F 2 , where the operators F i , i 1, 2 are defined as follows:
2.23
The following facts will be used in the proof.
x n − x 0,T 0, 2.28 for an x ∈ C 0, T , X .Then by the assumptions, we know that for almost every t ∈ 0, T and t, s ∈ Δ:
2.34
By 2.33 and our assumptions, we see that F is continuous.
Since χ is the Hausdorff measure of noncompactness in X, we know that χ is monotone, nonsingular, invariant with respect to union with compact sets, algebraically semiadditive, and regular.This means that i for any Noting that for any ψ ∈ L 1 0, T , X , we have lim So, there exists a positive constant L such that
10 Journal of Applied Mathematics
For every bounded subset Ω ⊂ C 0, T , X , we define mod c Ω : lim
2.41
Then mod c Ω is the module of equicontinuity of Ω, and α is a measure of noncompactness in the space C 0, T , X with values in the cone R 2 .
2.42
By the assumptions and the continuity of T t in the uniform operator topology for t > 0, we get mod c F 1 Ω 0.
2.43
Clearly, f s, x s a x s ≤ μ s m * k * x 0,T .
2.45
It is not hard to see that the right-hand side of 2.45 tend to 0 as t 2 → t 1 .Thus, the set Combining with 2.43 , we have mod c FΩ 0, which implies mod c Ω 0 from 2.42 .Next, we show that Ψ Ω 0. It is easy to see that
2.46
For any t ∈ 0, T , we define
2.47
We consider the multifunction s ∈ 0, t G s : Obviously, G is integrable, that is, G admits a Bochner integrable selection g : 0, h → E, and
2.51
Therefore, since X is a separable Banach space, we know by 20, Theorem 4.2.3 that
2.53
Similarly, if we set is integrable and integrably bounded.Thus, we obtain the following estimate for a.e.s ∈ 0, t :
2.56
Journal of Applied Mathematics 13 Now, from 2.53 and 2.56 , it follows that where 0 < L < 1.Then by 2.42 , we get Ψ Ω 0. Hence α Ω 0, 0 .Thus, Ω is relatively compact due to the regularity property of α.This means that F is α-condensing.
Let us introduce in the space C 0, T , X the equivalent norm defined as x * sup t∈ 0,T e −Lt x t .
2.59
Next, we show that there exists some r > 0 such that FB r ⊂ B r .Suppose on the contrary that for each r > 0 there exist x r • ∈ B r , and some t ∈ 0, T such that Fx r t * > r.
From the assumptions, we have Moreover,
2.62
Dividing both sides of 2.62 by r, and taking r → ∞, we have qM Γ 1 q sup t∈ 0,T t 0 t − s q−1 μ s m * k * e −L t−s ds ≥ 1.
2.63
This is a contradiction.Hence for some positive number r, FB r ⊂ B r .According to the following known fact.
Let M be a bounded convex closed subset of E and F : M → M a α-condensing map.Then Fix F {x : x F x } is nonempty.we see that problem 1.2 has at least one mild solution.
Next, for c ∈ 0, 1 , we consider the following one-parameter family of maps:
2.64
We will demonstrate that the fixed point set of the family H, is a priori bounded.Indeed, let x ∈ Fix H, for t ∈ 0, T , we have x τ ds .
2.66
Noting that the H ölder inequality, we have
Journal of Applied Mathematics 15
We denote y t : sup s∈ 0,t x s .
2.69
Let t ∈ 0, t such that y t x t .Then, by 2.68 , we can see y s ds.
2.70
By a generalization of Gronwall's lemma for singular kernels 22, Lemma 7.1.1, we deduce that there exists a constant κ κ q such that
2.71
Hence, sup t∈ 0,T x t ≤ w.Now we consider a closed ball:
2.72
We take the radius R > 0 large enough to contain the set Fix H inside itself.Moreover, from the proof above, F : B R → C 0, T , X is α-condensing.Consequently, the following known fact implies our conclusion: Let V ⊂ E be a bounded open neighborhood of zero and F : V → E a α-condensing map satisfying the boundary condition: x / λF x , 2.73 for all x ∈ ∂V and 0 < λ ≤ 1.Then, Fix F is nonempty compact.
Example
In this section, let X L 2 0, π , we consider the following nonlocal Cauchy problem for an integrodifferential problem: where ∂ q t is the Caputo fractional partial derivative of order 0 < q < 1; ξ ∈ 0, π ; k > 0 is a constant to be specified later; 1, . . ., j are continuous functions and there exists a positive constant b such that generates an analytic semigroup and uniformly bounded semigroup {T t } t≥0 on X with T t L X ≤ 1.Therefore, 3.1 is a special case of 1.2 . | 2,403.8 | 2012-05-13T00:00:00.000 | [
"Mathematics"
] |
Orientation Estimation by Means of Extended Kalman Filter, Quaternions, and Charts
An orientation estimation algorithm is presented. This algorithm is based on the Extended Kalman Filter, and uses quaternions as the orientation descriptor. For the filter update, we use measurements from an Inertial Measurement Unit (IMU). The IMU consists in a triaxial angular rate sensor, and an also triaxial accelerometer. Quaternions describing orientations live in the unit sphere of R . Knowing that this space is a manifold, we can apply some basic concepts regarding these mathematical objects, and an algorithm that reminds the also called “Multiplicative Extended Kalman Filter” arises in a natural way. The algorithm is tested in a simulated experiment, and in a real one.
I. INTRODUCTION
K NOWLEDGE about the mechanical state of a system is necessary in many engineering fields.The orientation of the system is an important part of this mechanical state.Fields like robotics, virtual reality, or vehicle navigation among others, could require knowledge of the orientation of a system for tasks like: -Controlling an Unmanned Vehicle.
-Knowing the orientation of a camera in a scenario.
-Knowing the heading of a vehicle in a navigation system.
-Transforming measurements taken in the vehicle reference frame to an extern reference frame.The problem presents two main issues that need to be addressed: • We need to choose an orientation descriptor.
• We need to choose an estimation approach.The orientation of a system is understood as the rotation transformation that relates two reference frames: the one whose orientation we are interested in, and the reference frame of the system with respect to which we want to express such orientation.It only makes sense to speak about one orientation with respect to another system.Knowing that an orientation is a rotation transformation, our issue is to choose the most convenient parameterization for this rotation transformation.The most used parameterizations are the Euler angles, and their analogous, Tait-Bryan angles, rotation vectors, rotation matrices, and unit quaternions.A fairly complete survey of orientation representations is given in [1].Unit quaternions have properties that make them preferable against the other parameterizations.Namely: 1) There are no singularities (we avoid the "gimbal lock", that is present in Euler angles).2) They describe the orientation in a continuous way (unlike axis-angle representation).3) Motion equations are linear with quaternions.4) They are determined by 4 parameters (in contrast with a rotation matrix, that needs 9 parameters).Because of these properties, unit quaternions have been the most widely used orientation representation since the early 1980s [2], and we also use them in this work.
The orientation estimation problem has been addressed using several approaches.In [3] it is provided a survey of methods for orientation estimation by far more complete than could be given in this work, and it would not make sense to repeat it here.The Kalman Filter with its nonlinear versions is the protagonist.But there is a major issue: unit quaternion do not live in the Euclidean space, where the Kalman Filter is defined.This fact leads to a variety of approaches in the application of this formalism.In particular, the known as "Multiplicative Extended Kalman Filter" is the method of choice because of its accuracy, its relative simplicity, its computational efficiency, and for being flexible to incorporate a great variety of measurements.However, there seem to be some aspects of it that are still not well understood.
The objective of this paper is to explore a new view point for the Extended Kalman Filter applied to the estimation of orientations represented by unit quaternions.Its final form is very similar to that of the "Multiplicative Extended Kalman Filter", but it gets rid of the probably tricky definition of the "reset" operation, and it arises the introduction of a new update called "chart update".Not being that different the structure of these two algorithms, it would not be unreasonable to rename this MEKF as "Manifold Extended Kalman Filter".
The algorithm developed here is designed to take measurements from an Inertial Measurement Unit (IMU) which returns acceleration, and angular velocity measurements.Yet, this design is easily modifiable in order to adapt it to other type of sensors.
The MEKF has been tested in a simulated experiment, together with the known Madgwick algorithm [4].It also has been implemented in a real system, and tested with a commercial IMU.
The remainder of the paper is organized as follows.In Section II, we introduce the main properties of unit quaternions.In Section III, we introduce the basic concepts of manifold theory that will be used in the algorithm development.In Section IV, we review the motion equations and measurement equations.In Section V, we present the developed equations for the state prediction.In Section VI, we present the developed equations for the measurement prediction.In Section VII, we present the developed equations for the filter updates.Section VIII displays the experimental results.Finally, we expose the conclusions, and we picture the future work pathways in Section IX.II.QUATERNIONS Quaternions are hypercomplex numbers with three different imaginary units {i, j, k} , and can be expressed as q = q 0 + q 1 i + q 2 j + q 3 k . ( They can also be expressed in a vectorial form as Quaternion product is defined by the Hamilton axiom which produces the multiplication rule Quaternions describing rotations can be built with a unit vector that defines the rotation axis, q , and the angle of rotation, θ , through Having this form, they satisfy the restriction This means that quaternions describing orientations live in the unit sphere of R 4 .This space has dimension 3, although its elements are determined using 4 parameters.We will use basic concepts of manifold theory to handle this kind of space.
III. BASICS OF MANIFOLD THEORY When dealing with the Kalman filter, the distribution of a random variable, x , is encoded by its mean, x , and its covariance matrix, P , defined as This can be done when our random variables are elements of an Euclidean space.But when a random variable is an element of a manifold our covariance matrix definition could lose sense.This is our case, where the random variable q − q , does not describe an orientation.Then we need to redefine our covariance matrix in a different way, but we can not change the form of the definition of the covariance matrix if we want to use the Kalman filter, because this precise form is used in its derivation.We will solve this problem by defining our covariance matrix in a different space.But first we will review some previous definitions: a) Definition.Manifold: A n-manifold, M n , is a topological space in which each point is locally homeomorphic to the euclidean space, R n .This is, each point x ∈ M n has a neighborhood N ⊂ M n for which we can define a homeomorphism f : N → B n , with B n the unit ball of R n .b) Definition.Chart: A chart for a topological space, M , is a homeomorphism, ϕ , from an open subset, U ⊂ M , to an open subset of the Euclidean space, V ⊂ R n .This is, a chart is a function with ϕ a homeomorphism.Traditionally a chart is expressed as the pair (U, ϕ) .
A. The Set of Charts
Assume we know the expected value of our distribution of quaternions, q .In such case, we can express any unit quaternion as q = q * δ , with δ another unit quaternion (unit quaternions together with their multiplication rule form a group).And then, we can define the set of charts The set of charts (9), is used in [5], but this work does not talk about charts, and what we call "chart update" is not applied.
In each chart, ϕ q , the quaternion q is mapped to the origin.As the space deformation produced in the neighborhood of the origin is small, being the variance small, the distribution in each chart will be similar to the distribution in the manifold.
The inverse transformations for these charts are given by ϕ −1 q (e q ) = q * 1 4 + e q 2 2 e q . (10) B. The Transition Map a) Definition.Transition map: Given two charts (U α , ϕ α ) and (U β , ϕ β ) describing a manifold, with Having the set of charts defined by (9), and having two charts centered in quaternions p and q , and related by p = q * δ , then our transition map takes the form
IV. MOTION EQUATIONS AND MEASUREMENT EQUATIONS
The state of the system is defined by an orientation, encoded by a unit quaternion q , and by an angular velocity measured in the reference frame attached to our system, given by a vector ω .The unit quaternion defines a rotation transformation that determines the orientation of the system.This transformation relates vectors (denoted as v ) expressed in a reference frame attached to the solid whose state we want to describe, with the same vectors (denoted as v ) expressed in an inertial reference frame in which the gravity vector is expressed as g = (0, 0, −1) .Thus, our rotation transformation will yield Using a rotation matrix, And using our unit quaternion, where this time v = ( 0 v ) , and q * is the complex conjugate quaternion, that being q a unit quaternion, it is also its inverse.
A. Motion Equations
Knowing what our quaternion means, we can write the motion equations for the random variables that we use to describe the state of our system: where τ is the process noise, associated with the torque acting on the system, and its inertia tensor.
B. Measurement Equations
This work uses an IMU as information source.We can write the measurement equations that relate the random variables describing the state of our system, with the random variables describing the measurements of our sensors as follows: where r a t is the noise in the accelerometer measurement, r ω t is the noise in the gyroscope measurement, and a t are nongravitational accelerations.
V. STATE PREDICTION
In this section we expose the evolution equations used to perform the prediction of the expected value of the state, and of its covariance matrix.
A. Evolution of the Expected Value of the State
Taking the expected value in equations ( 15) and ( 16), assuming the random variables q(t) and q ω t (t) = t t-∆t τ (τ ) dτ to be independent, and the expected value of the process noise, q ω t (τ ) , to be constant when τ ∈ [t − ∆t, t) , our differential equations are transformed into other ones whose solutions are
B. Evolution of the State Covariance Matrix
Since we need a covariance matrix with a form like (7), we will define the covariance matrix of the state in the set of charts defined by (9).In particular, for each filter update we will have an expected value for the unit quaternion describing the orientation, q , and a covariance matrix defined in the Euclidean space, R 3 , whose points are related with those of the unit sphere of R 4 through the chart ϕ q .The origin of R 3 is mapped with the q quaternion, and points around the origin represent quaternions in the neighborhood of q .
1) Differential Equations for our Charts: This result, and its derivation, is totally inspired and is almost equal to that which appears in [2].Having the definition for our charts in (8) and (9), we can find the differential equations for the δ quaternion using the differential equations for q , (15), and for q .And having the differential equations for the δ quaternion, we can find the differential equations for a point e = 2 δ δ0 on the charts: Note that, by the definition of the charts, the vector of random variables e t is expressed in the chart centered in the quaternion q t .For each instant, t , we have a quaternion q t defining the chart for that time.Then the differential equation ( 21) defines the evolution of the vector e that "travels" between charts.
2) Differential Equations for the Covariance Matrix: We define the covariance matrix for the state of our system by Notice that we do not write "∆e t " .By the new definition of the covariance matrix, the term e t can be interpreted as a displacement from the q t quaternion, which is mapped to the origin of R 3 by the chart.Note also that being the covariance matrix symmetric, we do not need to find the evolution of all its terms.We just need to find the evolution of the terms P ee , P eω , and P ωω .
We are looking for an estimation of the covariance matrix in t , using the information in t − ∆t .This is, we want to get P t|t-∆t from P t-∆t|t-∆t .For P ωω it is easy to find this relation.Assuming the random variables ω t-∆t and q ω t to be independents, If we had a function e t = f (e t−∆t ) , we could replace in (22), and perhaps obtain a relation similar to (23).But we are not able to find a closed solution for (21).However, we can find a differential equation for P t using this differential equation for e t .Starting from ( 22), with After replacing (21), assuming that higher moments are negligible compared to second-order moments, and remembering our assumption of independence of the random variables q and q ω t , and therefore, of e and τ , ( 25) and ( 26) can be approximated by 3) Evolution Equations for the Covariance Matrix: We are dealing with a system of inhomogeneous linear matrix differential equations.Generally, a system of this type is untreatable, but in our case the equations are sufficiently decoupled to be able to find a solution.
Given a solution for P ωω , we can find an approximate solution for P eω .And with this solution, we can find an approximate solution for P ee .Denoting Ω = [ ω t ] × we can write with Q ω t = Q ω 0 t , being Q ω 0 a constant matrix representing the process noise covariance per time unit.
VI. MEASUREMENT PREDICTION
In this section we expose the measurement equations used to perform the prediction of the expected value of the measurement, and of its covariance matrix.
A. Expected Value of the Measurement 1) Expected Value of the Gyroscope Measurement: Taking the expected value on (18), (31) 2) Expected Value of the Accelerometer Measurement: Taking the expected value on (17), knowing that the g t vector does not change, and assuming that the non-gravitational accelerations affecting our system, a t , does not depend on its orientation, Using the unit quaternion describing the orientation of our system, this relation takes the form And if we use the rotation matrix constructed from this unit quaternion,
B. Covariance Matrix of the Measurement
The measurement is related to the state by means of the measurement equations: We can approximate linearly the relationship around the expected values, x t , and r t , using a Taylor series the Jacobian matrices evaluated on the expected value of the random variables.Then, our prediction equation of the measurement covariance matrix takes the form 1) Gyroscope Block: The measurement equation for the gyroscope is linear.Using (18) and ( 35) we obtain 2) Accelerometer Block: In order of being consistent with the Kalman filter formulation, the acceleration term, a t , should be part of the noise in the measurement, since if it were not so, it should be part of the state.Then, measurement noise in our Kalman filter has two components: • r t : the main measurement noise.This noise comes from the sensor.• a t : non-gravitational accelerations acting on the system.
These accelerations obstruct the measurement of the g vector.
Recalling that we express the covariance matrix of the state in R 3 , and knowing that doing q t = q t in the manifold, is equivalent to do e t = 0 in this space, we will have for the accelerometer measurement equation: , and after some calculus, 3) Measurement Covariance Matrix: Assuming independence of all random variables involved in the measure, our prediction equation for the measurement is where • Q a t is the covariance matrix of the random variable a t .• R a t is the covariance matrix that describes the noise in the accelerometer measurement, which is modeled by the random variable r a t .• R ω t is the covariance matrix describing the noise in the gyroscope measurement, which is modeled by the random variable r ω t .
VII. UPDATE
Although the original Kalman filter algorithm just requires the Kalman update, the fact that our covariance matrix is expressed in a chart makes necessary the computation of a second update.
A. Kalman Update
The Kalman update is performed in the space where the covariance matrix is defined.This is, we do not perform the Kalman update in the manifold, but in the chart.Given the estimate of the covariance matrix of the state, P t|t-∆t , and the estimate of the measurement covariance matrix, S t|t-∆t , the optimal Kalman gain is computed as Given the gain, we can update the state distribution in the usual way: The new state distribution will be expressed in the chart centered on x t|t−∆t .In this chart, e t|t-∆t = 0 , but e t|t = 0 .
B. Manifold Update
In order to find the quaternion corresponding to the updated vector e t|t expressed in the chart, we must reverse the function ϕ q t|t-∆t (e t|t ) making use of the equation (10):
C. Chart Update
After the Kalman update, the new state distribution is expressed in the chart centered on x t|t-∆t .We must update the covariance matrix expressing it in the chart centered on the updated state, x t|t , so that our information is expressed as at the beginning of the iteration.In order of achieve this objective, we must use the concept of transition maps, that for our charts take the form of (11).Being non-linear this relation, we need to find a linear approximation: e p (e q ) ≈ e p (e q ) + ∂ e p ∂e q eq=eq ( e q − e q ) .(43) After differentiating our transition map and evaluating in e t|t , having identified the charts ϕ p = ϕ q t|t and ϕ q = ϕ q t|t-∆t , we find out Then, our update equations for the charts are VIII.EXPERIMENTAL VALIDATION In this section we present results obtained from a simulated experiment, and a real one.
A. Simulated Experiment
Our simulated experiment consists on the definition of a path, the extraction of simulated measurements, the processing of this measurements, and the evaluation of the algorithm performance.Only knowing the real state, we are able to define some metrics to measure the performance of our algorithm.Finally, we display a comparison of the algorithm developed in this paper, which will be called Manifold Extended Kalman Filter (MEKF), and the currently popular algorithm developed by Madgwick [4], whose code can be found in [6].
1) Experiment Setup: For testing our algorithm, we can think in a simple and intuitive simulation.Let us imagine that we can freely roam the surface of a torus, which is a manifold whose space can be described in R 3 by The torus of our simulation will have R = 0.2m , and r = 0.05m .We can define a path in the torus using a third parameter to set the other two: We will use the path defined by v φ = 1 rad/s and v θ = 3 rad/s , and we will travel the path around the torus 3 times.This path can be seen in Figure 1.
Fig. 1.Path followed on the torus in our simulation.Reference frames that define the orientation of the IMU can be observed.
Accelerations occurring in the IMU can be calculated by differentiating twice in (48)-(50) with respect to the t parameter, resulting Now, we can define a reference frame for each point of the path.The axis of the reference frame will have the directions of the vectors After making the derivatives in (56), and choosing the direction of the vectors so that z points outward the torus surface, our rotation matrix relating a vector measured in the IMU reference frame, with the same vector measured in the external reference frame will be Matrix (57) can be expressed as the product of 3 known rotation matrices: Recognizing these matrices in (58), and using (5), we can find out the quaternion describing the orientation of our reference frame: Having (59) we can obtain the q quaternion: where And with the q quaternion we can use (15) to get the angular velocity: With all, we can generate a succession of states q r ω r , and for each state simulate a measurement a m t ω m t .After that, taking only the succession of measurements, we can make a succession of estimations about the state q e ω e using the orientation algorithms, and then compare our estimation with the known real state q r ω r .2) Error Definition: We will evaluate the performance of the algorithm through the definition of two errors: First, defining g r = (R r ) T (0, 0, −1) T as the gravity vector measured in the real reference frame attached to our system, and g e = (R e ) T (0, 0, −1) T as the gravity vector measured in the estimated reference frame, we define The e g error in (64) is defined as the angle between the vectors g r and g e .Being the lowest error the better, this gives us a measure of how well the algorithm estimates the direction of a vector for which we have directly related measurements.
Second, we define q r i , and q r f as the initial and final quaternions describing the real orientation of the system in our simulation; and q e i , and q e f as the initial and final quaternions describing the estimated orientation of the system.Then, defining ∆ r , and ∆ e as the quaternions describing the rotation transformations that relates the initials and finals orientations by q r f = q r i * ∆ r , and q e f = q e i * ∆ e , we define The e θ error in (65) is defined as the angle of the rotation defined by the δ θ quaternion, which satisfies ∆ e = ∆ r * δ θ .Being the lowest error the better, this gives us a measure of how well the algorithm estimates the whole orientation, including heading, for which we do not have directly related measurements.This second error definition seems unnecessarily complicated.We could think in something like e θ = 2 arccos (q r f ) * * q e f , but if we start the simulation with an unknown orientation for the algorithm, this definition would lead to a different quaternion from the starting one.The e θ error would have a bias because of the ignorance of the initial heading.Our e θ error definition is independent of this initial heading ignorance.
3) Setting the Algorithm Values: a) Initialization: The simulation starts with a known orientation state defined by b) Characterization of Process Noise: The following values have been established: Still, after some testing, we found that the algorithm behaves similarly with other configurations, provided that they are not disproportionate.A more worked up algorithm would introduce dynamical values for this variables.c) Characterization of Measurement Noise: The Kalman filter requires in order to produce an unbiased estimation.
For the covariance matrices we will set and we will compare how the error behaves as a function of the magnitude of the noise in the measurement.
4) Simulation Results:
We will place our simulation in 4 different scenarios, whose details are displayed in table I, and that have been chosen according to the current possibilities.
Results of Test 1 and 2 are not shown since the errors produced are too large.This suggests that both a good processor (small sampling time) as a good sensor (small variance) are required.
For Test 3 and 4 the time evolution of error measurements are plotted in Figures 2 -5.In Figures 3 and 5 we observe how the error becomes smaller as we improve our IMU (decreasing σ 2 ), while having a good processor (small ∆t).In Figures 2 and 4 we observe how the error becomes smaller as we improve our processor (decreasing ∆t), while having a good IMU (small σ 2 ).In Figure 2 we observe that the MEKF increases its accuracy in time, due to it adds information about the state in each update.The faster it updates (less ∆t) the faster it reaches convergence.On the other hand the Madgwick algorithm does not adds information, which implies that it can not "learn" about the past, and it does not increases its accuracy in time.We also observe that after reaching a certain ∆t no improvement in accuracy is seen.
In Figure 3 we can confirm the same observation made in the paragraph above.But we also note that it seems to be a limit in the algorithms accuracy as a function of the sensor noise.It could be an interesting appreciation because it could mean that beyond a certain sensor quality, there would not be an appreciable improvement in the estimation.
In Figure 4 we can notice that the errors tends to increase over time.It is best appreciated for the Madgwick algorithm, and for low update frequencies with the MEKF.Probably it is not appreciated for higher update frequencies because we have not waited enough.This happens because we have no reference for orientation in the plane perpendicular to the gravity vector.If we want to have a complete non-biased estimation of the orientation we should add measurements from additional sensors as a magnetometer, or a camera.
In Figure 5 we again see the same behavior noticed in the previous paragraph.We also repeat our observation about the limit of the estimation accuracy as a function of the sensor noise made two paragraphs above.
B. Real Experiment
Our real experiment consists on the visual inspection of the returned information by our algorithm, and the one returned by the algorithm implemented in a commercial IMU. 1) Test Bed: The algorithm has been implemented in a real system.It has been used a board containing a MPU6050 sensor.The processing is performed in the ATmega328P chip contained in an Arduino board.In this system the sampling time turns out to be about (74) Figure 6 shows the assembled system, consisting in the MPU6050 sensor, the Arduino UNO, and a MTi sensor of Xsens.2) Experiment Results: We have described a series of movements with the system.The movements have been carried out in four phases.The dynamics of each phase has been more aggressive than that of the previous phase.We have tried to finish with the same orientation with which the system began.Both have been saved the sensors measurements and the estimated states which are returned by the algorithms.In Figures 7 -9 these data are shown.In Figure 7 we can see that both sensor acceleration measurements are very similar.This makes us think that the misalignment between the two sensors is small.
In Figure 8 we note that both systems return a similar estimation of the orientation when the dynamics is not too aggressive.However, after some aggressive moves, the algorithm presented in this paper has a fast convergence.We also note the bias in the heading estimation of both algorithms when we look at q 2 quaternion component.The initial and final orientation should be the same, but we have no reference for orientation in the plane perpendicular to the gravity vector.
In Figure 9 we note that the measured angular velocity is very similar to the estimated angular velocity.Perhaps we could accept the gyroscope measurement as the real angular velocity of our system.Maybe then we could get some advantage in processing speed, and therefore greater accuracy of our algorithm.But this is left for future research.
IX. CONCLUSIONS AND FUTURE WORK
We have successfully used basic concepts of manifold theory for estimating orientations using quaternions as descriptors.A similar algorithm to the known as "Multiplicative Extended Kalman Filter" naturally arises in applying these concepts without having to redefine any aspect of the Extended Kalman Filter.
The orientation has been estimated using measurements from an IMU, but the basic idea introduced in this work is applicable to any other type of sensor intended to estimate the orientation.
We have tested the algorithm in a real experiment and we have compared our estimation with the one given by a commercial IMU, finding that both orientation estimations are similar.This tell us that our algorithm works as expected.
We also have tested the algorithm in a simulation.We have compared the performance of our algorithm with the algorithm developed by Madgwick.The results suggest that the algorithm developed in this paper could achieve a better accuracy than the one achieved by the Madgwick algorithm.However we dare not say so, as there may be various sources of error that we have not considered: -The update frequency depends on the processor.In the ATmega328P chip, contained in an Arduino board, the Madgwick algorithm is 16 times faster than the MEKF.-The chosen path could lead to pathological behaviors because of its symmetry.-We have seen the result of just a path.These considerations lead us to the following issues that will be addressed in future work: • We will design a simulation with results based on the averaging of multiple trajectories, and free of pathological behaviors.• We will test the algorithm for various chart definitions.
• We will test its Unscented Kalman Filter version with the various chart definitions.• We will compare the MEKF with the MUKF, and with the Madgwick algorithm.
• We will study our sighting about the limit in the algorithm accuracy as a function of the sensor noise.
APPENDIX A STATE PREDICTION
In this appendix we present in greater detail the developments concerning the state prediction.
A. Evolution of the Expected Value of the State 1) Evolution in the Expected Value of the Angular Velocity: Taking the expected value in equation ( 16), 2) Evolution in the Expected Value of the Orientation: Taking the expected value in equation ( 15), Assuming the random variables q(τ ) and q ω t (τ ) to be independent,1 This differential equation has no general closed solution.But if we assume that the expected value of the process noise, q ω t (τ ) , is constant when τ ∈ [t − ∆t, t) , then we will have the matrix differential equation This differential equation has the solution q(t) = e Ω ∆t q(t − ∆t) .
We can express this solution using the quaternion product, as (79) 2) Differential Equations for the Covariance Matrix: a) Evolution Equation for P ωω : Assuming the random variables ω t-∆t and q ω t to be independents, their covariance is null.In such case, T × .
Here we can see the consequences of treating a nonlinear system.The evolution in the covariance matrix P ee , which is composed by moments of second order, is affected by the higher moments of the distribution.To find the evolution equations of the covariance matrix we would need information about the moments of order 3 and 4.These may depend of moments of order higher than them.Knowing all the moments of a distribution would mean to know all statistical information.What we can assume and expect is that higher moments to be negligible compared to second-order moments.
To find the solution to the inhomogeneous differential equation, we use the variation of constants method: Identifying the terms in the differential equation we obtain the following relation: To solve this last differential equation, it is necessary to propose a continuous evolution equation for P ωω .The simplest option is the linear function with Q ω t = Q ω 0 t , being Q ω 0 a constant matrix representing the process noise covariance per unit of time.
Having defined this continuous evolution equation, (88) is transformed into Integrating (90), The constant C is determined by the initial conditions P eω (0) = C(0) .With this in mind, Knowing that we have the information at t − ∆t , and we want to update the information at t , our equation becomes Finally, knowing that calculating infinite sums would take a lot, we can truncate in the first term, and write b) Evolution Equation for P ee : In this case we have the homogeneous equation Ṗee (τ ) ≈ − Ω P ee (τ ) − P ee (τ ) Ω T , whose solution is Using the variation of constants method, Identifying terms, we deduce the relation After substituting the expression for P eω , integrate with respect to time, and truncating in the first term of the infinite sums, the solution we want is given by T ∆t e −Ω T ∆t .
APPENDIX B MEASUREMENT PREDICTION
In this appendix we present in greater detail the developments concerning the measurement prediction.The g t vector does not change.If we also assume that the accelerations affecting our system, a t , does not depend on its orientation, One might be tempted to try to find the expected value of the matrix R t written as a function of the q t quaternion, but then we would run into the problem of the computation of expected values such as E[q 2 1 ] or E[q 1 q 2 ] .These expected values are defined in the manifold, and are what we try to avoid defining covariance matrices in the charts.
What we seek is not the expected value of the rotation matrix, but something like the "expected transformation".Then, using the quaternion describing the orientation of our system, this expression must be equivalent to Let us note that ∂ δt ∂et ∈ R 4×3 .However, the term APPENDIX C UPDATE In this appendix we present in greater detail the developments concerning the filter update.
Fig. 6 .
Fig. 6.Real system composed of an Arduino Uno, a MPU6050 chip, and a MTi sensor of Xsens.
Fig. 9 .
Fig. 9.Estimated and measured angular velocity during the real experiment.
A. Expected Value of the Measurement 1 )
Expected Value of the Accelerometer Measurement:Taking the expected value on (17), | 8,100.4 | 2017-07-04T00:00:00.000 | [
"Mathematics"
] |
Unified non-metric (1, 0) tensor-Einstein supergravity theories and (4, 0) supergravity in six dimensions
The ultrashort unitary (4, 0) supermultiplet of 6d superconformal algebra OSp(8∗|8) reduces to the CPT-self conjugate supermultiplet of 4d superconformal algebra SU(2, 2|8) that represents the fields of maximal N = 8 supergravity. The graviton in the (4, 0) multiplet is described by a mixed tensor gauge field which can not be identified with the standard metric in 6d. Furthermore the (4, 0) supermultiplet can be obtained as a double copy of (2, 0) conformal supermultiplet whose interacting theories are non-Lagrangian. It had been suggested that an interacting non-metric (4, 0) supergravity theory might describe the strongly coupled phase of 5d maximal supergravity. In this paper we study the implications of the existence of an interacting non-metric (4, 0) supergravity in 6d. The (4, 0) theory can be truncated to non-metric (1, 0) supergravity coupled to 5,8 and 14 self-dual tensor multiplets that reduce to three of the unified magical supergravity theories in d = 5. This implies that the three infinite families of unified N = 2, 5d Maxwell-Einstein supergravity theories (MESGTs) plus two sporadic ones must have uplifts to unified non-metric (1, 0) tensor Einstein supergravity theories (TESGT) in d = 6. These theories have non-compact global symmetry groups under which all the self-dual tensor fields including the gravitensor transform irreducibly. Four of these theories are uplifts of the magical supergravity theories whose scalar manifolds are symmetric spaces. The scalar manifolds of the other unified theories are not homogeneous spaces. We also discuss the exceptional field theoretic formulations of non-metric unified (1, 0) tensor-Einstein supergravity theories and conclude with speculations concerning the existence of higher dimensional non-metric supergravity theories that reduce to the (4, 0) theory in d = 6.
Introduction
Conformal supergravity theories with local Lagrangians based on the conformal superalgebras SU(2, 2|N ) have long been known to exist for N ≤ 4. It was generally believed that one could not go beyond N = 4 without having higher spins (> 2). In [1] it was shown that the fields of maximal N = 8 supergravity of Cremmer and Julia [2] can be fitted into an ultra short CPT-self-conjugate unitary supermultiplet of N = 8 superconformal algebra SU(2, 2|8) referred to as the doubleton supermultiplet. The corresponding ultra short supermultiplet of SU(2, 2|4) is the Yang-Mills supermultiplet in d = 4 [3]. The N = 4 Yang-Mills theory of doubleton supermultiplets of SU(2, 2|4) is conformally invariant both classically and quantum mechanically. This led the authors of [1] to pose the question whether a conformal supergravity theory based on the doubleton supermultiplet of SU(2, 2|8) exists which is closely related to the maximal N = 8 supergravity theory of Cremmer, Julia and Scherk. Since the latter theory is not conformally invariant any superconformal theory based on the doubleton supermultiplet of SU(2, 2|8) must be unconventional or exotic.
The superalgebra SU(2, 2|8) was used to classify the counterterms in maximal supergravity in [4]. Furthermore, it is known that amplitudes of maximal supergravity are SU (8) covariant even though the Lagrangian does not have SU (8) symmetry. This and above mentioned results provided part of the motivation for the work of Chiodaroli, Roiban and the current author [5] who studied the connection between maximal supergravity and superconformal symmetry in all dimensions that admit simple superconformal algebras as classified by Nahm [6]. They showed that the six dimensional counterpart of the doubleton supermultiplet of SU(2, 2|8) is the (4, 0) supermultiplet of the superconformal algebra OSp(8 * |8) with the even subalgebra SO * (8) ⊕ USp (8), where USp(8) is the R-symmetry group, which reduces to the CPT-self-conjugate doubleton supermultiplet of SU(2, 2|8) under dimensional reduction. They also showed that the (4, 0) theory can be obtained as a double copy of the (2, 0) theory based on the CPT-self-conjugate doubleton supermultiplet of OSp(8 * |4). 1 The (2, 0) supermultiplet first appeared in the work of [8] who constructed the entire Kaluza-Klein spectrum of 11-dimensional supergravity over AdS 7 × S 4 by simple tensoring of the (2, 0) doubleton supermultiplet. In the mid 1990s interacting (2, 0) supersymmetric theories in 6d were investigated within the framework of M/Superstring theory [9][10][11][12]. In particular, Seiberg pointed out the existence of four infinite series of new quantum theories with super-Poincare symmetry in six dimensions, which are not local quantum field theories [12]. Later an interacting (2, 0) superconformal theory was proposed by Maldacena as being dual to M-theory on AdS 7 × S 4 [13].
The (4, 0) supermultiplet was studied earlier by Hull using the formalism of double gravitons whose equivalence to the (4, 0) supermultiplet obtained using the twistorial oscillators was shown in [5]. Hull argued that an interacting (4, 0) theory in d = 6 might arise as the effective theory of the strongly coupled phase of five dimensional maximal supergravity when one of the dimensions decompactifies [14][15][16]. On the other hand the interacting (2, 0) JHEP06(2021)081 theory in six dimensions is believed to describe the strong coupling limit of 5d maximal super Yang-Mills theory. Since the maximal supergravity can be obtained as double copy of maximal super Yang-Mills theory in 5d these two proposals are consistent with the result that (4, 0) theory can also be obtained as double copy of (2, 0) theory in 6d [5,7]. More recently, the action for the free (4, 0) theory was written down by Henneaux, Lekeu and Leonard using the formalism of prepotentials in [17] based on their earlier work on (2,2) mixed chiral tensor describing the graviton [18]. The most unorthodox property of the (4, 0) doubleton supermultiplet of OSp(8 * |8) is the fact that the field strength of the graviton does not arise from a metric and hence the corresponding theory in 6d is sometimes referred to as non-metric , exotic or generalized supergravity. However under dimensional reduction it reduces to the standard maximal supergravity in five and four dimensions.
Independently of the work on maximal supergravity, five dimensional N = 2 supergravity theories coupled to vector multiplets (MESGT) were constructed in [19][20][21] and their gaugings were studied in [22][23][24][25][26][27]. Among these MESGTs four are very special in the sense that they are unified theories with symmetric scalar manifolds G/H such that G is a symmetry of the Lagrangian. They were called magical supergravity theories since their symmetry groups in five , four and three dimensions coincide with the symmetry groups of the famous Magic Square of Freudenthal, Rosenfeld and Tits [19]. Later it was shown that there exist three infinite families of unified MESGTs and two isolated ones [28]. Three of the magical supergravities belong to the three infinite families. The scalar manifolds of unified MESGTs outside the magical ones are not homogeneous. One infinite family of unified MESGTs can be gauged to obtain an infinite family of unified Yang-Mills Einstein supergravity theories in d = 5 with the gauge group SU(N, 1) [28].
In this paper we study some of the implications of the existence of an interacting 6d, non-metric (4, 0) supergravity theory. We show that the (4, 0) supergravity can be truncated consistently to non-metric (1, 0) supergravity coupled to 14, 8 and 5 self-dual tensor multiplets such that the resulting non-metric tensor-Einstein supergravity theories are unified theories in the sense that all the tensor fields including the gravitensor transform irreducibly under a simple global symmetry group. This in turn implies that all the three infinite families of unified 5d MESGTs as well as the two sporadic ones must also admit uplifts to non-metric unified (1,0) tensor-Einstein supergravity theories in d = 6. We conclude with some speculations about the possible extensions of the non-metric supergravity theories to higher dimensions with exotic spacetime signatures and the role of generalized superconformal algebras of these spacetimes.
The plan of the paper is as follows. In section 2 we review the 5d , N = 2 Maxwell-Einstein supergravity theories and their gaugings. Section 3 reviews the truncations of 5d, N = 8 supergravity to three of the magical supergravity and the symmetries of octonionic magical supergravity which can not be obtained from maximal supergravity. Section 4 reviews the uplifts of magical supergravity theories to six dimensions as Poincare supergravities. In subsection 5.1 we review the on-shell superfield formulation of (4, 0) supermultiplet and the gauge potentials in the "first order formalism" following [5] and give the gauge potential of the graviton field strength in the "second order formalism". In subsection 5.2 we review the exceptional field theoretic formulation of linearized (4, 0) JHEP06(2021)081 supergravity following the recent work of [29]. Subsection 5.3 is devoted to the question whether interacting conformal supergravity theories with SU(2, 2|8) symmetry in d = 4 and OSp(8 * |8) symmetry in d = 6 exist. In section 6 we give the truncations of (4, 0) supergravity to non-metric (3, 0) supergravity, to non-metric (2, 0) supergravity coupled to (2, 0) tensor multiplets and to non-metric (1, 0) supergravity coupled to (1, 0) tensor multiplets. In section 7 we discuss the metric and non-metric (1, 0) magical supergravity theories. Section 8 is devoted general unified non-metric (1, 0) tensor-Einstein supergravity theories in six dimensions and their exceptional field theoretic formulation. In section 9 we speculate about the possible extensions of non-metric supergravity theories to higher dimensions with non-standard space-time signatures. Appendix A reproduces the CPT-self-conjugate doubleton supermultiplet of SU(2, 2|8) [1].
2 Review of 5D, N = 2 Maxwell-Einstein supergravity theories and their gaugings N = 2 MESGTs in five dimensions describes the coupling of N = 2 supergravity to an arbitrary number, n, of vector multiplets. The supergravity multiplet consists of the fünfbein e m µ , two gravitini Ψ i µ (i = 1, 2) and one vector field A µ (the "bare graviphoton"). On the other hand a N = 2 vector multiplet consists of a vector field A µ , two symplectic Majorana spinor fields λ i and one real scalar field ϕ. The fermions in these theories transform as doublets under the R-symmetry group USp(2) R ∼ = SU(2) R while all the bosonic fields are SU(2) R singlets.
Hence the fields of an N = 2 MESGT can be labelled as where we labelled the bare graviphoton as A 0 µ . The indices a, b, . . . and x, y, . . . correspond to the flat and curved indices on the scalar manifold, M, respectively.
The bosonic part of the Lagrangian is given by [20] where e is the determinant of the fünfbein , R is the scalar curvature and F I µν are the field strengths of Abelian vector fields A I µ . The completely symmetric tensor C IJK , with lower indices is constant and determines the corresponding N = 2 MESGT uniquely [20]. The global symmetries of the Lagrangian JHEP06(2021)081 are the same as symmetries of C IJK . The n dimensional scalar manifold can be identified with a hypersurface in an (n + 1) dimensional ambient space with the metric with real variables h I (I = 0, 1, . . . , n) representing the coordinates of the ambient space. The scalar manifold M is simply the hypersurface V(h) = 1 and the metric , • a IJ (ϕ), of the kinetic energy term of vector fields is simply the restriction a IJ to M: The physical requirements of unitarity and positivity of the MESGT restrict the possible C-tensors. The most general C IJK that satisfy these constraints can be brought to the form where C ijk (i, j, k = 1, 2, . . . , n) are completely arbitrary. This is referred to as the canonical basis. Arbitrariness of C ijk implies that for a given number n of vector multiplets, there exist, in general, MESGTs with different scalar manifolds and different global symmetries.
Unified Maxwell-Einstein supergravity theories
Unified Maxwell-Einstein supergravity theories in d = 5 are those theories with a simple global symmetry group under which all the vector fields A I µ , including the graviphoton, form a single irreducible representation. With a combination of supersymmetry and global noncompact symmetry group any field can be transformed into any other field within this class of theories.
Among MESGTs whose scalar manifolds are homogeneous spaces only four are unified theories. They are defined by the four simple Euclidean Jordan algebras J A 3 of degree three defined by 3 × 3 Hermitian matrices over the four division algebras A, namely the real numbers R, complex numbers C, quaternions H and octonions O. The cubic norm defined by the C-tensor in these theories is identified with the cubic norm of the underlying Jordan algebra. They are referred to as magical supergravity theories because of the deep connection between their geometries and the geometries associated with the "magic square" of Freudenthal, Rosenfeld and Tits [30][31][32].
In N = 2 MESGTs defined by Euclidean Jordan algebras , J , of degree three the scalar manifold is a symmetric space of the form where Str 0 (J) and Aut(J) are the reduced structure and automorphism group of J, respectively. 2 Below we list the corresponding scalar manifolds: We should note that for MESGTs defined by Euclidean Jordan algebras of degree three such as the magical theories the C-tensor is an invariant tensor of the isometry group Str 0 (J) of the scalar manifold and we have where the indices I, J, K, . . . are raised by the inverse • a IJ (ϕ) of the metric of kinetic energy term of vector fields.
In addition to four unified MESGTs defined by four simple Euclidean Jordan algebras of degree three there exist three infinite families of unified theories whose scalar manifolds are not homogeneous as was shown in [28]. These three infinite families are defined by Lorentzian Jordan algebras of arbitrary degree. Now (n × n) Hermitian matrices over various division algebras form Euclidean Jordan algebras with the symmetric Jordan product defined as 1/2 the anticommutator. Their automorphism groups are compact groups. Non-compact analogs of these algebras, denoted as J A (q,n−q) , are generated by matrices over various division algebras , A = R, C, H for n ≥ 3 and over O for n ≤ 3, 3 that are Hermitian with respect to a non-Euclidean "metric" η with signature (q, n − q): It was shown in [28] that the structure constants (d-symbols) of traceless elements T I of noncompact Jordan algebras J A (1,N ) with Lorentzian metric η of signature (1, N ) defined as satisfy the unitarity and positivity requirements and can be identified with the C-tensor of a MESGT: The resulting MESGTs are all unified (for N ≥ 2 ) since all the vector fields including the graviphoton transform in a single irrep of the simple automorphism groups of the underlying Jordan algebras Aut(J A (1,N ) ) which are also the symmetry groups of their Lagrangians.
Unified N = 2 Yang-Mills-Einstein supergravity theories in five dimensions
A unified N = 2 Yang-Mills Einstein supergravity (YMESGT) theory is defined as a theory in which all the vector fields including the graviphoton transform in the adjoint representation of a simple non-Abelian subgroup of the global symmetry group that is gauged. Turning off the gauge coupling constant yields a unified MESGT under whose global symmetry group all the vectors transform irreducibly.
JHEP06(2021)081
In [28] the complete list of unified N = 2 YMESGTs in d = 5 was given. They are obtained by gauging the global SU(N, 1) symmetry groups of unified MESGTs defined by complex Lorentzian Jordan algebras J C (1,N ) under which all the vector fields transform in the adjoint representation of SU(N, 1). As stated above the MESGT defined by J C (1,3) is equivalent to the MESGT defined by the Euclidean Jordan algebra J H 3 whose global symmetry is SU * (6). Gauging the SU(3, 1) = SO * (6) subgroup of SU * (6) leads to the unique unified 5d YMESGT whose scalar manifold is a symmetric space [23]. Again in [23] it was shown that the dimensionless ratio g 3 κ involving the non-Abelian gauge coupling constant g and the gravitational constant κ must be quantized in the quantum theory by invariance under large gauge transformations. The same argument extends to all unified YMESGTs since where Π 5 stands for the fifth homotopy group.
Pure YMESGTs in d = 5 without tensor or hypermultiplets do not have a scalar potential. By expanding around the base point where a is some real number fixed by the condition d 000 = 1, one can show that the noncompact gauge fields transforming in N ⊕N of U(N ) become massive by eating scalar fields and around this ground state U(1) × SU(N ) remains unbroken with the U(1) gauge field corresponding to the graviphoton. Spin 1/2 fields transforming in the symplectic N ⊕N also become massive and together with massive gauge fields form short BPS multiplets, with the central charge generated by the U(1) factor.
N = 2 Yang-Mills-Einstein supergravity theories coupled to tensor fields
under the SU(N, 1) subgroup of USp(2N, 2) for N ≥ 2. Therefore in gauging the SU(N, 1) subgroup the N (N + 1) non-adjoint vector fields must be dualized to massive tensor fields satisfying odd dimensional self-duality conditions [25].
As for the family of unified MESGTs defined by the real Jordan algebras J R (1,N ) , the vector fields transform in the symmetric tensor representation of SO(N, 1). For even N = JHEP06(2021)081 2n with N > 3 one can gauge the U(n) subgroup of SO(2n, 1) by dualizing the non-adjoint vector fields transforming in the reducible symplectic representation n(n + 1) 2 ⊕ n(n + 1) 2 of U(n) to tensor fields. For odd N = 2n + 1 (N > 3) in gauging the U(n) subgroup of SO(2n + 1, 1) the remaining vector fields in the reducible representation of U(n) must be dualized to tensor fields.
In the MESGT defined by the octonionic Jordan algebra J O (2,1) with the global symmetry group F 4(−20) one can gauge the SU(2, 1) subgroup with the remaining vector fields transforming in the reducible representation of SU(2, 1) dualized to tensor fields.
Magical supergravity theories and maximal supergravity
The magical Maxwell-Einstein supergravity theories defined by the real, complex and quaternionic Jordan algebras J A 3 ( A = R, C, H ) can all be obtained by a consistent truncation of the maximal supergravity in d = 5, 4 and 3 dimensions [19]. The same is true for their 6 dimensional uplifts as Poincare supergravities [33]. The exceptional supergravity defined by the exceptional Jordan algebra J A 3 on the other hand can not be obtained by a truncation of maximal supergravity. In five dimensions the U-duality group of maximnal supergravity is E 6(6) and that of exceptional supergravity is E 6(−26) . They can both be truncated to the N = 2 MESGT defined by the quaternionic Jordan algebra with the U-duality group SU * (6). Maximal supergravity can be gauged in d = 5 with the gauge group SU(3, 1) and 12 tensor fields which admits an N = 2 supersymmetric vacuum with vanishing cosmological constant [34]. Similarly the exceptional supergravity theory can be gauged with the gauge group SU(3, 1) and 12 tensor fields. The common sector of these two gauged supergravity theories is the unique unified N = 2 YMESGT with the gauge group SU(3, 1) and whose scalar manifold is SU * (6)/USp (6).
JHEP06(2021)081
Under the maximal compact subgroup USp(6) the above representations of SU * (6) decompose as Under the maximal compact subgroup USp(8) of E 6(6) we have the decompositions On the other hand maximal compact subgroup of E 6(−26) is F 4 under which we have the decompositions Under the USp(6) × USp(2) subgroup the above representations of F 4 decompose as 26 = (14, 1) ⊕ (6, 2) (3.14) The above decompositions show that restricting to the USp(2) invariant subsector the spectra coincide with that of quaternionic magical theory defined by J H 3 . The global symmetry group SU * (6) of the quaternionic magical theory has the subgroup SL(3, C) × SO(2) where SL(3, C) is the global symmetry group of the complex magical MESGT The U(1) C invariant sector of the quaternionic theory corresponds to the consistent truncation to the complex magical theory. Similarly the global symmetry group of the complex magical theory decomposes as and Z 2 invariant subsector describes the consistent truncation to the real magical supergravity defined by J R 3 .
Magical Poincare supergravity theories in six dimensions
Six dimensional magical supergravity theories coupled to hypermultiplets and their gaugings were studied in [33]. Magical supergravities in six dimensions describe the coupling of (1, 0) Poincare supergravity to n T = 2, 3, 5, 9 tensor fields and vector fields in a definite spinor representation of SO(n T , 1). The coupling between vector fields and tensors involve SO(n T , 1) invariant tensors Γ I AB that are the Dirac Γ-matrices for n T = 2, 3, and the Van der Waerden symbols for n T = 5, 9. They satisfy the identities which are simply the Fierz identities for the existence supersymmetric Yang-Mills theories in 3,4,6 and 10 dimensions. These identities follow from the adjoint identity satisfied by the elements of simple Euclidean Jordan algebras of degree three [35]. We reproduce their field contents in table 2.
Since the vector fields transform in a spinor representation which belong to a unique orbit of the isometry group of the scalar manifold one finds that six dimensional magical supergravity theories admit a unique gauge group which is a centrally extended Abelian nilpotent group. For the octonionic magical theory the unique gauge group is the maximal centrally extended Abelian subgroup of F 4(−20) which is the automorphism group of the Lorentzian octonionic Jordan algebra J O (2,1) . For the quaternionic ( complex) magical theory the unique gauge group is the maximal centrally extended Abelian subgroup of USp(4, 2) ( SU(2, 1) ) which is the automorphism group of the Lorentzian quaternionic ( complex) Jordan algebra J H (2,1) (J C (2,1) ) . They satisfy the inclusions
JHEP06(2021)081
These results show that semisimple gaugings of the 5d magical supergravity theories do not admit uplifts to six dimensions as standard Lagrangian Poincare supergravities. On the other hand it is known that the 5d , N = 4 super Yang-Mills theory can be obtained from (1, 1) Poincare supersymmetric Yang-Mills theory in six dimensions by dimensional reduction and it is generally believed that it can also be obtained from an interacting (2,0) superconformal field theory. Similarly the N = 2 super Yang-Mills theory in five dimensions can be obtained from (1, 0) supersymmetric Yang-Mills theory or from a (1, 0) superconformal theory of self-dual tensor multiplets in d = 6. The standard Yang-Mills theories in d = 6 involving vector fields are not conformally invariant. The (2, 0) conformal supermultiplet decomposes as a (1, 0) tensor multiplet plus a conformal hypermultiplet in d = 6 Therefore an interacting (2, 0) theories can be viewed as a special family of interacting (1, 0) tensor multiplets coupled to hypermultiplets. Similarly the N = 4 super Yang-Mills theories that descend from the interacting (2, 0) theories in d = 6 can be viewed as a special class of N = 2 super Yang-Mills theories coupled to N = 2 hypermultiplets in the adjoint representation of the gauge group.
On-shell superfields of 6d , (4, 0) supermultiplet of OSp(8 * |8) in twistorial formulation and first versus second order formalism
The physical degrees of freedom corresponding to the fields of maximal N = 8 supergravity in d = 4 were shown to belong to the CPT-self-conjugate unitary representation (doubleton) of the conformal superalgebra SU(2, 2|8) in [1]. Formulation of this unitary supermultiplet in terms of constrained on-shell superfields was given in [5] which we review in the appendix. Even though the physical degrees of freedom form a unitary supermultiplet of the conformal superalgebra SU(2, 2|8) interactions in maximal supergravity break the conformal symmetry down to Poincare subgroup. Whether a conformal supergravity based on this supermultiplet exists, as contemplated in [1], is still an open problem as discussed below.
In six dimensions the unique superconformal algebra with 64 supersymmetry generators is OSp(8 * |8) with the maximal even subalgebra SO * (8)⊕USp (8). Explicit construction of the CPT self-conjugate doubleton supermultiplet of OSp(8 * |8) in terms of twistorial oscillators was given in [5] and shown to reduce to the doubleton supermultiplet of SU(2, 2|8) under dimensional reduction. In table 3 we reproduce the doubleton supermultiplet of OSp(8 * |8) with R-symmetry group USp(8) given in [5]. 4 This multiplet is referred to as the (4, 0) conformal supermultiplet in d = 6 and was studied earlier by Hull using the JHEP06(2021)081 formalism of double gravitons who argued that an interacting theory based on this supermultiplet may describe a strongly coupled phase of 5d maximal supergravity when one of the dimensions decompactifies [14][15][16].
The fields belonging to the (4, 0) supermultiplet can be fitted into an on-shell superfield satisfying an algebraic and a differential constraint [5]. For this it turns out to be very convenient to represent the coordinates of the six-dimensional extended superspace as antisymmetric tensors in spinorial indices [36,37] where the spinorial indices of the Lorentz group SU * (4) in d = 6 are labelled by hatted Greek indicesα,β . . . and the USp(8) indices by A, B, C . . . . Defining the superspace covariant derivative The symplectic metric satisfies and is used to raise or lower indices, The scalar superfield of the (4, 0) supermultiplet is completely anti-symmetric in its indices and is symplectic traceless i.e.
JHEP06(2021)081 and satisfies the differential constraint The mapping between twistorial formulation of (4, 0) supermultiplet and formulation in terms of vectorial indices M, N, . . . = 0, 1, . . . , 5 of SO(5, 1) as was done by Hull was given in [5]. The field strength Rαβγδ of the non-metric graviton corresponds to the (3, 3) which satisfies self-duality conditions in the first as well as the last 3 indices where * operation is performed with the Levi-Civita tensor in six dimensional Minkowskian spacetime In [5] the gauge potential for the non-metric graviton field strength Rαβγδ transforming in the (4, 0, 0) D representation of SU * (4) was chosen as a tensor field Cα (βγδ) transforming in the (3, 0, 1) D representation such that the field strength involves a single derivative where
It is invariant under the gauge transformations
Cα βγδ → Cα βγδ + ∂ω (β χωα γδ) (5.12) where the gauge parameters χαβ γδ satisfy However, since the standard Riemann tensor involves two derivatives of the metric one can also choose a gauge potential such that the non-metric graviton field strength involves two derivatives of that gauge potential as was done in [15,17]. In terms of spinorial indices such a gauge potential must transform as a mixed tensor Cαβ γδ satisfying the conditions Cαβ γδ = Cβα γδ = Cαβ δγ (5.14) such that the non-metric graviton field strength is given by which is invariant under the gauge transformations where the gauge parameters satisfy which imply that they transform in the 64 dimensional representation of SU * (4) with Dynkin label (1, 1, 1) D . The formulation given in [5] and the formulation in terms of a gauge potential of the form 5.16 given in [15,17] are the analogs of first and second order formalisms in ordinary supergravity [38]. 5 The underlying unitary (4, 0) supermultiplet that describes the physical degrees of freedom is the same for both formulations just as is the case for first and second order formalism of Poincaré supergravity. The gauge potential of the non-metric gravitino field strength is a traceless tensor ψα (βγ) . Under a gauge transformation it transforms as with the gauge parameter χ [αβ] γ such that χαβ β = 0. We should note that in contrast to standart local supersymmetry gauge parameter which involves a single spinor index, the gauge parameter χαβ γ of this local gauge symmetry of the non-metric gravitino field transforms as a spinor vector under SU * (4).
In terms of vectorial indices the non-metric gravitino field can be written as ψ [39,40] from those of N = 4 super Yang-Mills theory that is conformally invariant [41]. 6 Existence of an interacting 4d Poincare supergravity based on the conformal ultrashort CPT self-conjugate supermultiplet of SU(2, 2|8) suggests that an interacting Poincare supergravity in 6d based on the conformal (4, 0) supermultiplet of OSp(8 * |8) also exists in which the interactions break the superconformal symmetry OSp(8 * |8) down to its Poincare subgroup. On the other hand the existence of conformal supergravity based on the (4, 0) supermultiplet in six dimensions is an open question just like the existence of an interacting N = 8 conformal supergravity in d = 4 as will be discussed below.
In six dimensions the (2, 0) supermultiplet of OSp(8 * |4) is the conformal analog of 4d N = 4 Yang-Mills supermultiplet [8]. It is generally believed that the interacting theories of (2, 0) supermultiplets in six dimensions are not conventional field theories and may only exist as quantum theories [43]. Nonetheless they reduce to conventional field theories in lower dimensions. In [5] it was pointed out that the (4, 0) supermultiplet can be obtained by tensoring the (2, 0) supermultiplets. This raises the possibility that the "amplitudes" or correlation functions of an interacting superconformal (2, 0) theory could yield the amplitudes of an interacting non-metric Poincare supergravity based on the (4, 0) supermultiplet via some generalization of double copy methods of BCJ.
Even though the supermultiplet of fields of (4, 0) supergravity and their free equations were known for a long time the action for linearized non-metric (4, 0) Poincare supergravity in six dimensions was first written down rather recently in [17]. The authors of [17] use the formalism of prepotentials adapted to the self-duality properties of the fields of the (4, 0) supermultiplet. They show that the resulting action is invariant under (4, 0) Poincaré supersymmetry in d = 6 but not manifestly. The reason for loss of manifest Poincaré covariance is due to the fact that to write down the action they split the 6d spacetime coordinates as 5+1 with the singlet coordinate being timelike. In their Lagrangian formulations of the bosonic self-dual tensors which they refer to as chiral two-forms [44] as well as bosonic chiral (2,2)tensor corresponding to the gauge potential of non-metric graviton [18] involve only spatial tensors, and their temporal components are pure gauge. 7 They also present a similar formulation for the non-metric gravitensorino ("chiral spinorial two-form"). The resulting action of free (4, 0) supergravity in terms of prepotentials is fourth order in spatial derivatives. 8 (4, 0) supergravity as well as the (3, 1) supergravity were also studied within the socalled exceptional field theory (ExFT) formalism by the authors of [29] recently. Before giving the exceptional field theoretic formulation of these theories they first present novel JHEP06(2021)081 actions for the bosonic sectors of linearized (4, 0) and (3, 1) using the 5 + 1 split of six dimensional spacetime coordinates such that the singlet coordinate y is space-like. These actions are two-derivative actions that reduce to the bosonic sector of linearized maximal supergravity in five dimensions.
The exceptional field theory formalism is a particular extension of the double field theory formalism and is an outgrowth of the attempts to make the hidden U-duality groups of lower dimensional supergravity theories manifest in higher dimensions from which they can be obtained by toroidal compactification. 9 To achieve this one intoduces an auxiliary "internal" spacetime with coordinates Y M motivated by the U-duality group with which to extend the standard external d-dimensional spacetime with coordinates x µ and imposes a section constraint such that the resulting theory describes the higher dimensional supergravity theory. For 5d maximal supergravity one introduces a 27 dimensional auxiliary internal space-time extending the 5 dimensional external spacetime and imposes a section constraint of the form where C IJK is the E 6(6) invariant symmetric tensor. The above equation is to be interpreted as differentials acting on functions To recover 11-d supergravity corresponding to the 5 + 6 split of the coordinates one decomposes the 27 internal coordinates Y I with respect to SL(6, R) × GL(1) subgroup of E 6(6) 27 = 6 +1 + 15 0 + 6 −1 ⇔ Y I = (y m , y mn = −y nm ,ȳ m ) (5.25) where m, n, . . . = 1, . . . , 6. By restricting the dependence on Y I only to the 6 coordinates y m one obtains a solution to section constraints that leads to the 11 dimensional supergravity [46]. 10 To construct the linearized 6d (4, 0) , (3, 1) and standard (2, 2) supergravity theories as ExFTs in a unified manner describing the maximal N = 8 supergravity in 5d the authors of [29] extend the 27 dimensional internal space-time with an extra singlet coordinate Y • and impose the more general section constraint where ∆ IJ is a constant tensor describing the background spacetime. Setting ∂ • = 0 one has the standart section constraint of the formulation of the maximal N = 8 supergravity as an ExFT whose solutions include the 11d sugra, type IIB supergravity and maximal (2, 2) Poincare supergravity theory in 6d. On the other hand setting ∂ I = 0 the generalized section constraint 5.26 is satisfied trivially and by identifying the extra coordinate Y • with 9 For a review and references on the subject see [45]. 10 Type IIB supergravity corresponds to the decomposition of 27 of E 6(6) with respect to the SL(5, R) × SL(2, R×GL(1) subgroup given by 27 = (5, 1) −4 +(5 , 2) −1 +(10, 1) −2 +(1, 2) − 5 and restricting dependence on internal coordinates to the coordinates y a in (5, 1).
JHEP06(2021)081
the space-like coordinate in the (5+1) split of 6d coordinates x µ they show that one obtains the bosonic sector of linearized (4, 0) supergravity. In addition to (2, 2) and linearized (4, 0) supergravity in 6d the generalized section constraint admits a solution corresponding to (3,1) supergravity at the linearized level as well. They leave to future work the supersymmetric extension of the bosonic sector of (4, 0) supergravity. Furthermore the fact that a unified framework exists for ExFT formulations of 6d (2,2), (4,0) and (3,1) supergravity theories and the (2, 2) theory can be extended to the full nonlinear theory is interpreted by the authors of [29] as evidence that the same may be true for the (4, 0) and (3, 1) theories. If the interacting non-metric (4, 0) supergravity exists as a Lagrangian theory it will admit a formulation as an ExFT describing the uplift of maximal Poincare supergravity to six dimensions. However if the interacting (4, 0) theory is non-Lagrangian we shall assume that an appropriate generalization of the ExFT formalism exists with which to uplift the 5d maximal supergravity to 6d as a non-metric (4, 0) Poincare supersymmetric theory.
On the question of existence of interacting N = 8 conformal supergravity theory in d = 4 and (4, 0) conformal supergravity in d = 6
The existence of an interacting conformal supergravity based on the (4, 0) supermultiplet in d = 6 would suggest the existence of an interacting conformal supergravity in d = 4 based on the CPT self-conjugate doubleton supermultiplet of SU(2, 2|8) whose existence was posed as an open problem in [3]. To this date no such supergravity theory has been constructed. Conformal supergravity theories were first studied in the pioneering papers of [47,48]. 11 Recently they have been studied as massless limits of Einstein-Weyl supergravity theories. 12 The standard conformal supergravity theories in d = 4 based on the conformal superalgebras SU(2, 2|N ) exist only for N ≤ 4. All N =4 conformal supergravities in d = 4 have recently been constructed in [51,52]. In addition to a massless graviton they contain a massive spin two ghost field and hence are not unitary. The constraint N ≤ 4 arises from the fact that for N ≥ 4 the conformal supermultiplets containing the massive spin two field must necessarily contain fields of spin greater that two. Furthermore the vector fields associated with the gauge fields of SU(n) and U(1) subgroups inside U(n) ⊂ SU(2, 2|n) have kinetic energy terms that are of opposite sign and hence are not all positive definite. Hence if an interacting N = 8 superconformal theory exists that is unitary its formulation must go beyond the standard local gauging of the underlying conformal superalgebras. It may exist purely at the quantum level without a Lagrangian formulation or its Lagrangian formulation may be non-local. In fact there are non-local formulations of conformal gravity that are both unitary and ultraviolet finite. 13 Their Lagrangians typically involve infinite number of terms that are bilinear in scalar curvature R, Ricci tensor R µν and Riemann tensor R µνρλ with powers of the D'Alembertian sandwiched between them. These results are consistent with the findings of [55] who studied the four 11 For an older review see [49]. 12 For a review see [50] and the references therein. 13 For reviews we refer to [53,54] and references therein.
JHEP06(2021)081
derivative action of Stelle [56] of the form where M P l is the Planck mass, ∆ is the cosmological constant and µ indicates the mass scale. This theory has a massive spin two ghost of mass µ whose kinetic energy has opposite sign to that of the massless mode. They show that there exists a ghost free completion of this theory which requires an infinite series of higher derivative terms. The resulting theory is classically equivalent to ghost-free bimetric theory of two symmetric tensor fields studied in [57]. Whether these ultraviolet finite unitary nonlocal (higher derivative) theories admit supersymmetric extensions that go beyond the N = 4 bound and result in a unitary theory whose massless physical spectrum coincides with the CPT-self-conjugate doubleton supermultiplet of SU(2, 2|8) is an open problem. The question about the existence of interacting conformal supergravity theory with OSp(8 * |8) symmetry in d = 6 is more subtle. Such a theory would necessarily involve a nonmetric "graviton" and no interacting non-metric gravity theories have been constructed to date. The metric conformal supergravity theories in d = 6 with superconformal symmetry OSp(8 * |2N ) exist only for N ≤ 2). Whether there exist non-local ( higher derivative) and non-metric conformal supergravity theories with OSp(8 * |8) symmetry and are unitary is an open problem.
Truncations of (4, 0) supergravity
Assuming that there exists an interacting (4, 0) supergravity that reduces to the maximal N = 8 supergravity one can consider its truncations to interacting theories with lower number of supersymmetries.
JHEP06(2021)081
The (2, 0) supersymmetric truncation of the (4, 0) conformal supergravity reduces to N = 4 supergravity coupled to 5 vector multiplets in d = 5 with the global symmetry group SO(5, 5) × SO(1, 1) which is also the global symmetry group of the 6 dimensional theory with the moduli space The tensor fields of the six dimensional theory form the (10+1) representation of SO(5, 5)× SO (1, 1). The resulting supergravity can be gauged in d = 5 to obtain Yang-Mills Einstein supergravity theories with various possible gauge groups, in particular SU(2) × U(1). Since it is generally believed that strongly coupled phase of 5d maximal super Yang-Mills is described by an interacting (2, 0) theory in d = 6 this raises the question whether there exist deformations of the non-linear (4, 0) theory that correspond to various gaugings of maximal supergravity in d = 5 or in d = 4. In general not all gaugings of maximal supergravity are expected to have uplifts to higher dimensions since gaugings in general introduce potentials with anti-de Sitter as well as de Sitter vacua.
transforming as symplectic traceless anti-symmetric tensor of rank three which we will denote as 14 .
The (1, 0) truncation of (4, 0) theory that reduces to the quaternionic magical theory in 5d can be further truncated such that the resulting theory describes a unified theory in d = 5. First, by restricting to the U(1) invariant sector in the decomposition of SU * (6) with respect to its subgroup SL(3, C) × U(1), which requires discarding 6 tensor multiplets, one obtains a theory describing the coupling of non-metric (1, 0) graviton supermultiplet to eight tensor multiplets that reduces to the complex magical supergravity theory in d = 5 whose moduli space is SL(3, C)/SU (3). Second, by a further restriction to Z 2 invariant sector under the decomposition SL(3, C) ⊃ SL(3, R)×Z 2 one obtains a 6d theory describing the coupling to 5 tensor multiplets that reduces to the real magical supergravity in 5d with the moduli space SL(3, R)/SO (3).
In all the above truncations of the interacting (4, 0) theory to a (1, 0) supersymmetric theory describing the coupling of generalized graviton multiplet to tensor multiplets all the tensor fields, including the gravitensor, transform in an irrep of the global symmetry group, which are SU * (6), SL(3, C) and SL(3, R), respectively. The quaternionic theory with global SU * (6) symmetry can be extended to the octonionic theory with the global symmetry group E 6(−26) symmetry by coupling additional 12 tensor multiplets and reduces to the octonionic magical supergravity theory with 27 vector fields in d = 5.
(3, 0) supersymmetric truncation of (4, 0) supergravity
Using the labelling of indices as in the previous subsection one can show that by discarding two N = 6 gravitensorino multiplets consisting of the fields one obtains the non-metric (3, 0) graviton supermultiplet consisting of the fields The truncation to the non-metric (3, 0) supergravity theory has the same bosonic field content as the truncation to the maximal (1, 0) tensor Einstein supergravity with 14 selfdual tensor multiplets and the 14 scalars. The corresponding result for the 5d supergravity, namely that N = 6 supergravity has the same bosonic content as the quaternionic magical theory with the symmetric target space SU * (6)/USp (6) as was shown in [19].
ExFT formulation of metric and non-metric (1, 0) magical supergravity theories in six dimensions
The unified ExFT formulation of the bosonic sector of linearized (4, 0) and (2, 2) Poincare supergravities as formulated in [29] can be readily extended to a unified construction of JHEP06(2021)081 (1, 0) metric and non-metric supergravity theories in six dimensions that descend to the four magical supergravity theories in five dimensions. For the octonionic magical supergravity which can not be obtained from maximal supergravity by truncation one imposes the section constraint where the C-tensor C IJK is the one given by the cubic norm of the real exceptional Jordan algebra J O 3 which is Euclidean. It an invariant tensor of E 6(−26) . To obtain the ExFT formulation of the metric (1, 0) supergravity in d = 6 one first decomposes the indices of 27 dimensional representation of E 6(−26) with respect ot its SO(9, 1) subgroup where η ab is the SO(9, 1) invariant metric and (Γ a ) αβ are the gamma matrices of SO(9, 1) with (a, b, . . . = 0, 1, . . . 9) and (α, β, . . . = 1, 2, . . . , 16). Imposing the conditions solves the section constraint and the resulting ExFT describes the metric octonionic magical supergravity in d = 6 with 9 tensor multiplets and 16 vector multiplets and scalar manifold SO(9, 1)/SO(9) [33]. For constructing the non-metric (1, 0) supergravity theory one imposes the condition ∂ I = 0 and identifies the coordinate Y • with the spatial singlet component y in the (5 + 1) split of external 6d spacetime coordinates x µ as was done for the maximal supergravity in [29].
To obtain the unified ExFT formulations of metric and non-metric (1, 0) supergravity theories that reduce to the linearized quaternionic, complex and magical supergravity theories in d = 5 one needs to simply substitute the C-tensors of these theories in the section constraint 5.26 and decompose the indices with respect to subgroups of their 5d U-duality groups listed in the first column of table 2 and the gamma matrices by those listed in column 4 of that table. For the quaternionic magical theory the tensor C IJK becomes the invariant tensor of the internal Lorentz ( reduced structure) group SU * (6) of J H 3 . The resulting metric (1, 0) supergravity describes the coupling of 6 tensor multiplets and eight vector multiplets to metric (1,0) supergravity. The corresponding non-metric (1, 0) supergravity theory describes the coupling of 14 tensor multiplets to non-metric (1, 0) supergravity. In contrast to the metric theory the non-metric supergravity theory describes a unified theory since the 15 tensor multiplets transform irreducibly under the global symmetry group SU * (6). Since the quaternionic magical supergravity can be embedded both in maximal supergravity and octonionic magical supergravity in five dimensions the corresponding ExFts describing the metric and non-metric (1, 0) supergravity theories in 6d can also be obtained by truncation of the unified formulation of ExFTs describing (4, 0) and (2, 2) theories in 6d.
The ExFT defined by J H 3 can be consistently truncated to a (1, 0) unified non-metric tensor-Einstein supergravity described by the Euclidean Jordan algebra J C 3 describing the JHEP06(2021)081 coupling of 8 tensor multiplet to non-metric (1, 0) supergravity. It is an invariant tensor of the Lorentz ( reduced structure ) group SL(3, C) of J C 3 . The latter theory can be further truncated to an ExFT corresponding to the real magical supergravity in d = 5 defined by J R 3 . Its C-tensor is invariant under SL(3, R) and describes the coupling of 5 self-dual tensors to non-metric (1,0) supergravity in six dimensions.
The enlarged symmetry groups SL(3, C), SU * (6) and E 6(−26) are the global symmetry groups of the 5d MESGTs defined by the corresponding Euclidean Jordan algebras. As we discussed above these three 5d unified MESGTs theories can be obtained from unified tensor-Einstein supergravity theories in 6d. For the non-metic tensor-Einstein supergravity theories to be unified theories all the tensor fields including the gravitensor must transform irreducibly under the global symmetry group. Remarkably the corresponding irreducible representations of SL(3, C), SU * (6) and E 6(−26) remain irreducible under the restriction to the manifest symmetry subgroups SO(3, 1), SU(3, 1) and USp(6, 2), respectively. For the other members of the three infinite families of unified MESGTs there is no symmetry enhancement beyond the automorphism groups of the underlying Lorentzian Jordan algebras. Nonetheless we expect them to descend from unified tensor-Einstein supergravity theories in 6d in a similar fashion. In addition to the 3 infinite families there exist a unified MESGT in d = 5 defined by the Lorentzian octonionic Jordan algebra of degree three J O (1,2) . This isolated theory is also expected to descend from a unified tensor-Einstein supergravity in d = 6. In tables 4 and 5 we give the list of unified tensor-Einstein supergravity theories in d = 6, their field content and global symmetry groups, under which all the tensor fields including the gravitensor form an irrep.
As is evident from the tables 4 and 5 in the decomposition of the irreducible representation of the global symmetry group with respect to its maximal compact subgroup there is a unique singlet which is to be identified with the "bare gravitensor". Furthermore all the bosonic fields in tensor Einstein supergravity theories are singlets of the R-symmetry group USp(2).
JHEP06(2021)081
One can gauge certain subgroup of the global symmetry groups of the unified MESGTs in d = 5. This naturally leads to the question whether the corresponding six dimensional tensor-Einstein supergravity theories admit interactions among tensor fields that reduce to the non-Abelian gauge interactions in five dimensions. It is generally believed that the interacting superconformal (2, 0) theories do not admit a Lagrangian formulation since they decribe multiple M-5 branes that are strongly coupled with no free parameter for formulating a perturbative Lagrangian theory. Nonetheless a novel method of introducing such non-Abelian couplings in certain (1, 0) superconformal field theories in d = 6 was developed by Samtleben and collaborators [58][59][60]. It is an open problem whether some of these theories can be coupled to non-metric (1, 0) supergravity in 6d that upon dimensional reduction reduce to 5d, N = 2 Yang-Mills-Einstein supergravity theories. (4, 0) and (1, 0) tensor Einstein supergravity theories
Dimensional reduction of non-metric
Dimensional reduction of (4, 0) supergravity multiplet using vectorial indices was performed by Hull who showed that the resulting field content coincides with that of maximal Poincare supergravity in five dimensions. Dimensional reduction of the (4, 0) unitary supermultiplet using the twistorial spinor indices was given in [5]. Twistorial oscillator method yields manifestly unitary supermultiplets which involve only the physical degrees of freedom. Hence the resulting supermultiplets involve only the field strengths and not the gauge fields. For the doubleton supermultiplet of OSp(8 * |8) of the (4, 0) non-metric supergravity in d = 6 the field strengths correspond to symmetric tensors in the spinor indicesα,β, . . . of the Lorentz group SU * (4). Since the spinor representation Sα of 6 dimensional Lorentz group SU * (4) and symmetric tensor representations Sαβδ ,... remain irreducible under the restriction to the five dimensional Lorentz group USp(2, 2), the dimensional reduction to d = 5 in the twistorial formulation as given in [5] is much simpler than in the formulation involving vectorial indices. There is a one-to-one correspondence between the field strengths belonging to the doubleton supermutiplet of OSp(8 * |8) and the field strengths of the fields of maximal supergravity in d = 5. In particular non-metric graviton in 6d with the field strength R (αβδγ) reduces to the five dimensional graviton field strength without an extra vector or scalar field [5,15]. This is to be contrasted with the 6d metric graviton which reduces to a graviton plus a vector and a scalar. Similarly the gravitensor field strength ψ A (αβγ) remains irreducible under restriction to USp(2, 2) subgroup of SU * (4) and becomes the field strength of a gravitino in 5d: where * γ µν = 1 3! µνλ δ γ λ δ . Self-dual tensor fields reduce to vector fields in d = 5 and the symplectic Majorana Weyl spinors go over to symplectic Majorana spinors in d = 5. Therefore non-metric (1, 0) supersymmetric tensor Einstein supergravity theory in d = 6 will reduce to a five dimensional N = 2 supergravity with the same number of 5d vector fields as self-dual tensors in 6d. The bare gravitensor in 6d will reduce to the bare graviphoton in 5d. What distinguishes unified tensor-Einstein supergravity theories from JHEP06(2021)081 others is the fact that gravity sector can not be decoupled from the tensorial matter sector without breaking their global symmetry groups since the gravitensor together with the other tensor fields transform irreducibly under them. Ungauged tensor Einstein theories reduce to Maxwell-Einstein supergravity theories in d = 5.
Unified MESGT theories, in particular the magical supergravity theories , admit gaugings with simple gauge groups with or without tensor fields in five dimensions. It was shown in [33] that Poincare uplifts of magical supergravity theories to six dimensions do not admit gaugings with simple groups. Furthermore Poincare uplifts of magical supergravity theories in 6d are no longer unified since some of the vector fields uplift to selfdual tensors while the others uplift to vector fields in 6d. In addition one finds that magical Poincare supergravity theories in 6d admit a unique gauge group which is a nilpotent Abelian group with (n T −1) translation generators, where n T is the number of tensor multiplets coupled to (1, 0) metric supergravity. These (n t − 1) generators can not lie within the isometry group SO(n T , 1) of the scalar manifold due to appearance of central extensions of the gauge algebra. The gauge algebra with the central extension can be embedded within the isometry group of the corresponding five dimensional magical supergravity. On the other hand in the uplift of the magical supergravity theories to six dimensions as non-metric (1, 0) tensor Einstein supergravity the isometry group of the scalar manifold of the five dimensional theory becomes a global symmetry group in 6d.
On the general ExFT formulation of unified non-metric tensor-Einstein supergravity theories in six dimensions
The 27 dimensional internal space-time that is intoduced in ExFT formulation of the 5d maximal Poincare supergravity can be identified with the generalized spacetime coordinatized by the split exceptional Jordan algebra J Os 3 of 3 × 3 Hermitian matrices over the split octonions O s . This generalized space-time was first intoduced in the early days of spacetime supersymmetry before any supergravity theories was written down [61]. Its automorphism, reduced structure and linear fractional groups were identified with the rotation, Lorentz and conformal groups of this space-time which are F 4(4) , E 6(6) and E 7(7) respectively. The generalized space-times defined by Jordan algebras were later studied further in [62][63][64][65]. For the maximal supergravity in d = 5 the Lorentz group E 6(6) is the invariance of the C-tensor as well as of the Lagrangian.
For the 5d octonionic magical supergravity theory the internal space-time is the generalized spacetime defined by the real exceptional Jordan algebra J O 3 of 3 × 3 Hermitian matrices over the division algebra of octonions whose rotation, Lorentz and conformal groups are F 4 , E 6(−26) and E 7(−25) , respectively. The invariance groups of the C-tensors of magical supergravity theories are given by the Lorentz groups of underlying Euclidean Jordan algebras of degree three. As was discussed above the complex, quaternionic and octonionic magical supergravity theories can equivalently be described by the Lorentzian Jordan algebras of degree four over the reals R , complex numbers C and quaternions H. MESGTs in 5d defined by Lorentzian Jordan algebras J A (1,n) of arbitrary degree over the reals R , complex numbers C and quaternions H.
The maximal (2, 2) supergravity theory in six dimensions can be truncated to (1, 0) supersymmetric Poincare supergravity that reduces to the magical supergravity defined by the Euclidean Jordan algebra J H 3 in five dimensions. Its global symmetry group in six dimensions SU * (4) × USp(2) which is a subgroup of the5d global symmetry group SU * (6).
and describes the coupling of 5 self-dual and 8 vector multiplets to (1, 0) metric supergravity. The global symmetry group of this theory gets enlarged to SU * (6) in five dimensions with the scalars parametrizing the symmetric space SU * (6)/USp (6). This theory can be further truncated to the complex and real magical supergravity theories. The dimensional reduction of unified ExFT formulation of linearized metric and non-metric magical supergravity theories to five dimensions the resulting theories are invariant only under the maximal compact subgroups of their global symmetry groups which are F 4 , USp(6), SU(3) and SO (3).
As summarized in section 2.1 5d MESGTs described by the Euclidean Jordan algebras J H 3 and J C 3 can be equivalently formulated using the Lorentzian Jordan algebras J C (1,3) and J R (1,3) , respectively. Since the section constraint for a unified ExFT formulation of these theories depends only on the C-tensor it extends to the formulation in terms of Lorentzian Jordan algebras. Interestingly the extra singlet coordinate Y • in the ExFT formalism can now be identified with the identity element of the corresponding Lorentzian Jordan algebra. Therefore for the three infinite families as well as the sporadic unified non-metric (1, 0) TESGTs in d = 6 one can give an ExFT formulation of their linearized bosonic sectors using the section constraint where C IJK are now the structure constants of the underlying Lorentzian Jordan algebras which are invariant tensors of their global symmetry groups given by their automorphism groups. In all cases the extra singlet coordinate Y • can be identified with the identity element of the underlying Lorentzian Jordan algebra. Imposing the condition ∂ I = 0 solves JHEP06(2021)081 the section constraint trivially and leads to the bosonic sector of non-metric (1, 0) TESGTs in d = 6. Apart from the magical supergravity theories the higher dimensional origins of the three infinite families of 5d unified MESGTs as Poincare supergravity theories is not known. Whether the unified section constraint admits solutions that lead to three infinite families of (1, 0) supersymmetric metric Poincare supergravities in six dimensions will be left for future investigations.
Conformal path to higher dimensions and non-metric supergravity theories
If an interacting non-metric (4, 0) supergravity exists in d = 6 that reduces to the maximal supergravity in lower dimensions it raises the question as to whether d = 6 is the maximal dimension for the existence of non-metric supergravity theories. Now six is the maximal dimension for the existence of conformal superalgebras that extend the conformal algebra of SO(d, 2) to a simple conformal superalgebra [6]. Here the isomorphism of SO(6, 2) to SO * (8) plays a key role for satisfying spin and statistics constraints. The extensions of the Lie algebra of SO(d, 2) to simple Lie superalgebras for d > 6 do not satisfy the correct spin and statistics connection. Therefore it is believed that conformal metric supergravity theories based on simple superconformal algebras exist only in d ≤ 6.
On the other hand Poincare superalgebras, which are not simple, exist in any dimension. However maximal dimension for Poincare supergravity is d = 11 [6]. The Poincare superalgebra in d = 11 with 32 supercharges can be embedded in the simple Lie superalgebra OSp(1|32, R) with the even subalgebra Sp(32, R) [66][67][68][69]. Its contraction to 11 dimensional Poincare superalgebra involves tensorial central charges Werner Nahm's classification of spacetime superalgebras assumed that the spacetime has Minkowskian signature and gravity is described by a spacetime metric or a corresponding spin connection. In the (4, 0) supermultiplet of OSp(8 * |8) the gauge field of the graviton field strength is not a spacetime metric but rather a mixed tensor both in the first order as well as the second order formalism as reviewed above. As was shown by Hull the gauge symmetries of the mixed tensor reduce to diffeomorphisms in five dimensions and the interacting theory should yield the standard maximal supergravity in five dimensions. This raises the question whether there could be higher dimensional non-metric supergravity theories which reduce to the interacting (4, 0) theory in d = 6 or standard maximal supergravity in five and lower dimensions. At the level of Lie superalgebras the answer appears to be yes. Namely there exist superalgebras of the form OSp(2n * |2m) which extend the
In this context we should point out that M-theory was studied in space-times with exotic signatures by Hull [70]. Later M-theory and superstring theory in exotic signature space-times were studied within the framework of negative branes in [71]. Typically the Lorentz groups of these exotic spacetimes are of the form SO(p, 11−p) or SO(p−10−p) for formulations of M-theory and IIB superstring theory, respectively and the corresponding gravitational theories are of metric type. To our knowledge formulation of M/superstring theory on exotic spacetimes with conformal groups of type SO * (2n) that admit interpretation as non-metric gravity theories has not yet been investigated.
One natural framework for going beyond standard Minkowskian spacetimes with metric gravity is that of generalized spacetimes coordinatized by Euclidean Jordan algebras [61,63,64,72]. The conformal groups of space-times defined by Euclidean Jordan algebras all admit positive energy unitary representations [63] and were shown to be causal spacetimes in [65]. The standart critical Minkowskian space-times that admit supersymmetric Yang-Mills theories can be coordinatized by Euclidean Jordan algebras J A 2 of degree two generated by Hermitian 2 × 2 matrices over the four division algebras A = R, C, H, O. Their conformal groups are Sp(4, R), SU(2, 2), SO * (8) and SO(10, 2), respectively. Except for the octonionic case they all admit extensions to simple Lie superalgebras, namely OSp(n|4, R), SU(2, 2|n) and OSp(8 * |2n) which all satisfy the usual spin and statistics connection. 14 The natural extension of these space-times is to consider those defined by Euclidean Jordan algebras J A 3 of degree 3 over the four division algebras A = R, C, H, O which were studied in [64]. These spacetimes correspond to extensions of the Minkowskian space-times by twistorial coordinates given in the second column of table 2 and an extra singlet coordinate. The C-tensors given by the norm forms of Jordan algebras of degree three satisfy the so-called adjoint identity C IJK C J(M N C P Q)K = δ I (M C N P Q) (9.2) It is a remarkable but little known fact that for simple Jordan algebras of degree three the adjoint identity implies the Fierz identities for the existence of supersymmetric Yang-Mills theories in the critical dimensions [35]. The conformal groups of J A 3 are Sp(6, R), SU(3, 3), SO * (12) and E 7(−25) , respectively. Again except for the octonionic case they admit extensions to simple superalgebras OSp(n|6, R), SU(3, 3|n) and OSp(12 * |2n), respectively. The conformal superalgebras in critical dimensions 3, 4 and 6 are subalgebras of these superalgebras.
We should note that SO * (4n) is the conformal group of a spacetime coordinatized by the Euclidean Jordan algebra J H n of n × n Hermitian matrices over the division algebra of quaternions [63]. Its Lorentz and rotation groups are SU * (2n) and USp(2n), respectively. The quaternionic Jordan algebra J H 2 of degree two describes the standard 6d Minkowskian space-time with the Lorentz group SU * (4) and conformal group SO * (8) , which are isomorphic to Spin(5, 1) and SO(6, 2), respectively. The conformal superalgebras OSp(8 * |2n) JHEP06(2021)081 ( 1 2 , 0) no conformal supergravity with the same field content as maximal N = 8 supergravity in four dimensions has been constructed. It is clear that such a supergravity would have to have some unusual properties such as non-locality and higher derivaties. On the other hand the superalgebra SU(2, 2|8) and the above supermultiplet play a key role in the construction and classification of potential higher-loop counterterms in maximal supergravity [4]. The physical fields of the CPT-self-conjugate doubleton supermultiplet of SU(2, 2|8) can be represented as a scalar superfield W abcd [5,75,76] in the superspace with coordinates, where α,β = 1, 2 and a = 1, 2, . . . 8. The covariant derivatives in this superspace satisfy The superfield W abcd is completely anti-symmetric and obeys the self-duality condition as well as the differential constraint | 14,786.8 | 2021-06-01T00:00:00.000 | [
"Physics"
] |
Non-Hermitian BCS-BEC crossover of Dirac fermions
We investigate chiral symmetry breaking in a model of Dirac fermions with a complexified coupling constant whose imaginary part represents dissipation. We introduce a chiral chemical potential and observe that for real coupling a relativistic BCS-BEC crossover is realized. We solve the model in the mean-field approximation and construct the phase diagram as a function of the complex coupling. It is found that the dynamical mass increases under dissipation, although the chiral symmetry gets restored if dissipation exceeds a threshold.
Introduction
One of the fundamental tenets of modern quantum mechanics and quantum field theory is that Hamiltonians should be Hermitian. It ensures that energy levels are real and the time evolution is unitary. However, it has been perceived in recent years that there are situations in which physics can be effectively described in terms of non-Hermitian Hamiltonians [1,2]. Such a description proves to be useful for open quantum systems that interact with environments [3][4][5]. In PT -symmetric quantum mechanics, the existence of PT symmetry ensures real energy spectra even when the Hamiltonian is not Hermitian [6][7][8]. This is not just a mathematical possibility but can be realized experimentally [9]. It is not an overstatement to say that the physics of non-Hermitian Hamiltonians is far richer than conventional Hermitian ones; in fact, symmetry classification [10,11] reveals that there are 38 distinct symmetry classes in non-Hermitian systems, while there are only 10 in Hermitian systems. There are a plethora of exotic phenomena that are unique to non-Hermitian systems, such as the non-Hermitian skin effect [12][13][14][15][16] and the non-Hermitian localization transitions [17].
Superfluidity and superconductivity are amongst the most salient and fascinating phenomena in quantum many-body physics [62,63]. Non-Hermitian fermionic superfluidity has been explored in [64][65][66][67][68]. 1 In [67] the authors pointed out that ultracold atomic gases with two-body losses due to inelastic collisions can be naturally described with a complexvalued interaction. They solved the gap equation for fermions on a square/cubic lattice and mapped out the phase diagram in the mean-field approximation. In [68] this analysis was extended to fermions in a continuum model, where the phase diagram across the entire range from weak to strong coupling was obtained. This is a non-Hermitian analog of the well-established BCS-BEC crossover of fermions with s-wave interactions [62,[80][81][82][83][84][85]]when the s-wave scattering length is varied, the system evolves continuously from a weakly interacting BCS regime of loose Cooper pairs to the BEC regime of tightly bound molecules.
The primary goal of the present paper is to generalize the analysis of [67,68] to Dirac fermions and investigate a non-Hermitian relativistic BCS-BEC crossover. 2 What is the motivation of this study? First, Dirac fermions can be realized with ultracold atoms loaded on an optical lattice [86][87][88] and hence the experimental protocol proposed in [67] to produce a complex coupling can in principle be applied to this case as well. Second, a relativistic BCS-BEC crossover is believed to take place in QCD and QCD-like theories at finite density [89][90][91][92][93][94][95][96][97][98] (see [99] for a review). While the interaction strength and the density can be separately varied in nonrelativistic systems, this is not the case in QCD -actually we face a density-induced crossover: at low density, quark matter is strongly coupled and tightly bound diquarks condense, whereas at high density the coupling is weak due to asymptotic freedom and a BCS-type description is justified [100]. The study of such a crossover (including the possibility of phase transitions at intermediate densities) is potentially relevant to compact star phenomenology and heavy-ion collision experiments.
In this paper we investigate χSB in the NJL model with a complex coupling. Generally, χSB is considered to be a strong coupling phenomenon and a weakly coupled BCS picture does not apply. However, at finite chiral chemical potential [21,[101][102][103][104][105][106][107][108], χSB occurs at an arbitrarily weak coupling due to the fact that a chiral chemical potential induces a nonzero density of states for fermions at low energy and serves as a catalyst of χSB [107,108]. By tuning both the coupling strength and the chiral chemical potential, we probe the entire range of the BCS-BEC crossover for χSB and construct a complete phase diagram as a function of the complex four-fermion coupling. A novel mechanism for emergence and disappearance of complex saddles of the action is also illustrated.
2 Although [49] solved a zero-dimensional model of Dirac fermions with a complex four-fermion coupling, its higher-dimensional analog has not been thoroughly studied yet.
a chirally imbalanced matter in gauge theories is intrinsically unstable due to the axial anomaly [109][110][111], our model has no coupling to gauge fields and there is no instability due to anomalies. This paper is organized as follows. In section 2 the model is defined and the thermodynamic potential is derived. In section 3 the phase diagram for real coupling is presented. The detrimental effect of the baryon chemical potential on χSB is illustrated. In section 4 we turn on an imaginary part of the coupling, solve the gap equation numerically, map out the phase diagram, and determine the boundary between the normal phase, a metastable χSB phase, and a stable χSB phase. The results are then compared with those for nonrelativistic fermions [67,68]. We conclude in section 5.
The NJL model
We consider a model with the partition function where the Euclidean Lagrangian is given by µ is the quark chemical potential and µ 5 is the chiral (or axial) chemical potential. We set the current mass to zero. Eq. (2.1) is the same model as in [107,108] except that here we have nonzero µ.
It is straightforward to introduce N f flavors of quarks and let N f → ∞ with G ∝ 1/N f , which allows us to rigorously justify a saddle point analysis, but we shall stick to N f = 1 for simplicity of exposition.
With the Hubbard-Stratonovich transformation, quarks can be readily integrated out and yields The bosonic fields are related to fermionic observables as σ = G ψψ and π = G ψiγ 5 ψ .
In the mean-field approximation, we have where a momentum cutoff Λ was introduced to remove UV divergences. Let us define dimensionless variables which leads to a dimensionless action S ≡ S/Λ 4 given by In the following, we will assume 0 < µ 5 < 1 so that the Fermi surface stays inside the domain of integration. In generic open quantum systems, temperature is not well defined, and we treat t as a formal parameter used to define the path integral for the partition function, as in [67]. We will set t to zero in the ensuing analysis.
Phase diagram for real coupling
Let us begin with a discussion for real coupling g > 0. In the zero-temperature limit t → +0, (2.8) reduces to Numerical minimization of S allows us to determine the dynamical mass M as a function of g and µ 5 . Our result for µ = 0 is presented in Figure 1. As one can see from the left plot, while there is a nonzero critical coupling ≈ 20 at µ 5 = 0, it goes away for µ 5 > 0: χSB occurs for any nonzero coupling. This is the catalysis effect emphasized in e.g., [107,108]. The dynamical mass increases monotonically with µ 5 . What is the distinction between the BCS regime and the BEC regime? In nonrelativistic systems the unitarity limit marks a crisp boundary, but what about relativistic systems? The locus of a "transition" between the two regimes is not uniquely defined. As suggested in [99], one popular criterion is to see the dispersion relation E(p) of quasiparticles. If E(p) takes a minimum at |p| ≈ p F , it is in the BCS regime, and if E(p) is a monotonically increasing function of |p|, it is in the BEC regime. This works pretty well in dense QCD where E(p) = ( p 2 + M 2 − µ) 2 + ∆ 2 (with ∆ the superconducting gap) experiences such a transition when the dynamical mass M is equal to µ. In the current model, however, the dispersion (2.6) takes a minimum at |p| = µ 5 regardless of the dynamical mass. Yet another criterion is to compare the interparticle distance and the size of Cooper pairs. If the wave function of Cooper pairs extends beyond the average interparticle distance, it is in the BCS regime, otherwise in the BEC regime. In [113,114] the size of Cooper pairs in relativistic color superconductors was computed as a function of the quark density and such a BCS-BEC-type crossover was indicated. Unfortunately, a similar analysis is difficult for our model because the interaction is pointlike and the gap has no momentum dependence. As a rule of thumb, let us take the inverse of M as the size of a Cooper pair, and take the inverse of µ 5 as the average inter-quark distance. Then the region with 1/M > 1/ µ 5 (1/M < 1/ µ 5 ) corresponds to the BCS (BEC) regime, respectively. According to this crude estimate we labeled each regime in the right plot of Figure 1.
Next we proceed to the analysis for nonzero µ. The phase diagrams are shown in Figure 2. The main observation here is that for any µ 5 , there is a critical µ beyond which the chiral symmetry is restored. This is because µ induces a mismatch of Fermi surfaces and disrupts Cooper pairing. The critical value of µ is known as the Chandrasekhar-Clogston limit [115,116]. Analogous situations arise in both condensed matter [117,118] and QCD [119][120][121][122][123][124]. When the ordinary isotropic Cooper pairing is hampered, a nonstandard pairing that breaks translation symmetry is likely to set in, though it is beyond the scope of this paper.
Phase diagram for complex coupling
Finally we complexify the coupling constant. Throughout this section we set µ = 0 to simplify the ensuing numerical analysis. Eq. (2.8) reduces to where g ∈ C and M ∈ C. When M is purely imaginary, (x ± µ 5 ) 2 + M 2 may be exactly on the branch cut of the complex square root, which makes the zero-temperature limit t → +0 ill-defined. When M is not purely imaginary, we have for t → +0 This integral can be performed with the formula where tanh −1 (z) = 1 2 log 1+z 1−z is the inverse function of tanh(z). This way we obtain Since S is a function of M 2 , the trivial vacuum M = 0 is always a solution to ∂S/∂M = 0. We are interested in the gap equation for M = 0, which reads (4.6) We have varied g on the complex plane and numerically searched for a solution to (4.6) for each g. It turned out that there was no solution for Re g < 0, indicating that chiral symmetry is unbroken for Re g < 0. This is natural because Re g < 0 is a repulsive interaction. Furthermore, the phase structures for Im g > 0 and Im g < 0 are symmetric about the real axis of g. Hence we will assume Re g > 0 and Im g > 0 in the following. By monitoring the magnitude of the gradient ∂S/∂M we found an interesting mechanism that changes the number of saddle points of S. In Figure 3 Figure 3(b), a new saddle point is suddenly born out of the imaginary axis of M . So there are now two saddle points. When the imaginary part of g is further increased, as shown in Figure 3(c), the old saddle is absorbed into the imaginary axis and we are left with a single saddle. In this fashion the number of saddles (i.e., the solutions to (4.6)) can jump abruptly.
When there are multiple saddles, the dominant one is definitely the one that has the lowest value of Re S. Following [67,68], we define three phases as below.
• Normal phase: (4.6) has no solution, i.e., M = 0 is the only saddle of S. • Metastable χSB phase: there are solutions to (4.6), but their Re S are higher than that for M = 0. • Stable χSB phase: there are solutions to (4.6) whose Re S are lower than that for M = 0.
In Figure 4 (left) we display the phase diagram for µ 5 = 0.5 on the complex g plane. In the vicinity of the real axis we have a stable χSB phase. As Im g increases, we are driven into a metastable χSB phase via a quantum phase transition. At small Re g the metastable saddle goes away and the chiral symmetry is completely restored. Interestingly enough, the normal phase at small Re g sharply ends at Im g 22.5; this means that a very strong dissipation can trigger χSB (albeit a metastable one). The global phase structure we found here is quite similar to the one for nonrelativistic fermions on a lattice [67], though we note that the phase diagram at small Re g was not presented in [67] because of the limitation of numerical calculations.
To gain more insights, in Figure 4 (right) we plot the magnitude of the gap |M |. When there are multiple saddles, we took the one that has the lowest Re S. Notice that nothing dramatic happens at the boundary between the metastable χSB phase and the stable χSB phase. It is worth noting that the gap magnitude tends to be enhanced by dissipation. As a crude guide we drew the boundary |M | ∼ µ 5 between the BCS regime and the BEC regime. We emphasize that this is a rule of thumb and a more rigorous characterization of the crossover region in non-Hermitian superfluids is left as an open problem. We point out that for Im g 20 the gap reaches the UV cutoff scale (|M | ∼ 1); this implies that all the calculations for Im g 20 are extremely sensitive to the regularization scheme used, and hence one has to be careful about physical interpretations. Figure 5 shows the real and imaginary parts of M . The left panel shows that Re M grows monotonically with Re g and is largely independent of Im g. Note that Re M approaches zero along the boundary with the normal phase. This means that, when one moves out of the normal phase, a nontrivial solution to the gap equation emerges out of the imaginary axis of M . The right panel shows that Im M grows monotonically with Im g.
It is instructive to compare our findings with preceding works.
• Ref. [67] found an enhancement of superfluidity by dissipation, and attributed it to the continuous quantum Zeno effect which suppresses tunneling and reinforces onsite molecule formation. This effect is unique to a lattice system. Nevertheless we observed a similar enhancement in a continuum model; we speculate that the presence of a hard momentum cutoff plays a role similar to that of a lattice. • Ref. [68] solved the BCS-BEC crossover for a complex scattering length and found that the superfluid phase is stable even in the limit of strong dissipation, which is surprising. The difference of [68] from [67] and this paper may be due to the fact that [68] considered a complex-valued chemical potential. 3 • Ref. [66] reports that a non-Hermitian perturbation (an imaginary magnetic field) enhances superfluidity. • Ref. [34] solved the NJL model supplemented with a non-Hermitian bilinear term that preserves chiral symmetry. It was found that, as the non-Hermitian coupling grows, the dynamical mass first rises and then drops to zero.
Finally we turn to the quasiparticle excitation spectra E(p) = (|p|/Λ − µ 5 ) 2 + M 2 which is generally complex for complex g. The real and (scaled) imaginary parts of E(p) are displayed in Figure 6. At weak coupling, the imaginary part of E(p) is sharply concentrated around the Fermi level (left panel). By contrast, at stronger coupling, the imaginary part has a much broader support (right panel). These are results for weak dissipation (Im g = 0.5). If dissipation is stronger, a more peculiar thing can occur: at the boundary between the normal phase and the metastable χSB phase, M is purely imaginary (cf. Figure 5), hence E(p) is purely imaginary for µ 5 − |M | ≤ |p|/Λ ≤ µ 5 + |M | and is real otherwise. The bounds |p|/Λ = µ 5 ± |M | are an example of the so-called exceptional points [127] where the dimensionality of the eigenspace of the Hamiltonian decreases.
Conclusions and outlook
In summary, we have investigated χSB of Dirac fermions in four dimensions at finite chiral chemical potential for a complex-valued four-fermion coupling. We varied the coupling over the entire range from the weakly bound BCS regime to the tightly bound BEC regime, and numerically constructed the phase diagram. Our primary result is that the imaginary coupling tends to enhance χSB up to a certain threshold; when the imaginary coupling exceeds the threshold, the chiral symmetry is restored, although a nontrivial solution to the gap equation continues to exist. We illustrated how complex saddles of the action come into existence and go away through the imaginary axis of the complex gap plane along which the action has a branch cut. We also worked out the complex energy spectra of quasiparticles. Our results can in principle be tested in experiments using ultracold atomic gases and other materials that host Dirac fermions [128][129][130]. 4 However, it may be difficult to draw implications for quark matter in compact stars from this work, because (i) observable signals from compact stars are extremely scarce, and (ii) the pointlike four-fermion interaction is a very crude approximation to the non-Abelian gauge interaction between quarks.
There are miscellaneous future directions in which this work can be extended. A partial list is given below.
• To incorporate the competition between the chiral condensate and the diquark condensate (i.e., a competition between a Dirac mass and a Majorana mass) along the lines of [131,132] • To investigate collective fluctuations around the saddle points and test the validity of the mean-field approximation for non-Hermitian χSB • To analytically prove numerical findings in this work, e.g., the existence of a nontrivial solution to the gap equation for large Im g • To examine dissipative χSB under an external magnetic field that catalyzes χSB [133,134] • To study U(1) A vortices • To study the effect of nonzero µ on non-Hermitian χSB • To analyze the structure of Cooper pairs as a function of the complex coupling, and provide a more precise description of the non-Hermitian BCS-BEC crossover • To use the renormalization group to improve the mean-field analysis • To clarify how to apply the Lefschetz thimble approach [58] to non-Hermitian χSB where the complex action and its gradient have branch cuts • To use the results of this paper to benchmark algorithms (such as the complex Langevin method [135]) for simulating complex-action theories • To extend this work to lower spatial dimensions • To test the conclusions of this paper with other low-energy effective models such as holography [136] (see [137] for a non-Hermitian extension) | 4,478.6 | 2020-09-28T00:00:00.000 | [
"Physics"
] |
Behavioral Modeling of GaN Doherty Power Amplifiers Using Memoryless Polar Domain Functions and Deep Neural Networks
In this paper, novel Doherty Power Amplifier (DPA) models are presented. The motivation behind the proposed models is to accurately predict static nonlinearities in the compression regions of the carrier and peaking amplifiers. DPAs suffer from a nonlinearity that originates from the carrier amplifier, and a second more pronounced nonlinearity generated at the full compression region following the turn-on of the peaking amplifier. Moreover, these distortions are often observed at different input power levels depending on whether the AM-AM or the AM-PM characteristic is considered. Therefore, the proposed static model is based on independent modeling of the memoryless gain in the polar domain. The static model of the memoryless AM-AM and AM-PM characteristics is augmented with either memory polynomials or deep neural network functions for memory effects modeling. The methodology of building the proposed models and the achieved results are discussed in this paper. The MP based proposed model achieves an NMSE as low as −45.3dB with only 78 model parameters, while the DNN based model achieves an NMSE as low as −46.1dB with only 156 model parameters. However, the DNN based model achieves the best model resilience to changes in the identification data.
I. INTRODUCTION
The power amplifier (PA) is a major device included in a transceiver system. Its performance significantly impacts the quality of the transmitted signal and that of the communication link [1]- [4]. Modern PAs are driven by signals having various schemes such as Orthogonal Frequency Division Multiplexing (OFDM). With such advanced schemes, the static and memory effects of the PA should be included in the modeling and linearization processes [5]- [7]. There are many techniques and models developed for linearizing PAs. Each model has a different structure, complexity (number of model parameters) and error performance. Digital Pre-distorters (DPD) and power amplifiers (PA) can be modelled using memory polynomials [5]- [7] as well as neural The associate editor coordinating the review of this manuscript and approving it for publication was Rocco Giofrè . networks (NN) [8]- [16]. Real-valued and complex-valued NNs are developed and published in the literature. For example, a real-valued time-delay NN for modeling of 3G base-station PA was developed in [9]. The parameters of the NN were determined by backpropagation method. Different optimization methods can still be used such as particle swarm optimization and genetic algorithm coupled with local minimum search algorithm. The NN model can properly model PA characteristics with satisfactory agreement between simulated and measured data. Also, the dynamic AM-AM and AM-PM can be modelled using the real-valued NN to account for the memory effects. The advantage of using real-valued NNs over complex-valued NNs is that there is no need to use complex-valued backpropagation algorithm to determine the unknown parameters of the model [9]. Other forms of NNs published in the literature include Tapped Advance and Delay Line Neural Net (TADNN) [10]. TADNN models are VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ based on the feedforward tapped delay line structure. The importance of using the advance taps and the delay taps is to improve the performance of the DPD [10]. It was proven that the TADNN scheme improves PA linearization as compared to conventional delay-tap-only model. A second type of real-valued NN is the Two Layer Artificial Neural Network (2LANN) [11]. This model was developed for pico-cell PAs which output about 2 watts or less. The motivation behind this NN is to realize a model which can be efficiently implemented in a real-time application with hardware considerations. The 2LANN model provides acceptable performance with significant reduction in computational resources. A different approach for using NN in PA modelling is described in [12]. It is a real-valued autoregressive with exogenous input NN which can model a PA with memory effects. Complex-valued NN are also developed in the literature. A complex-valued two hidden layers NN (2HLANN) was published in [13]. Furthermore, a recurrent neural network termed the instant gated recurrent neural network (IGRNN) was developed in [14]. Bidirectional long-short-term memory networks (BiLSTM) have been introduced in the literature [15]. This architecture outperforms many current deep learning techniques for DPD and PA modeling especially when the PA exhibits strong memory effects. Furthermore, convolutional neural networks (ConvNet) have been also shown in the literature to be effective for modeling PA behavior [16]. The ConvNet model was found to reduce model complexity by more than 50% as compared to other deep learning techniques [16]. Other types of DPD and PA models are based on the memory polynomial models (MP). Unlike NN models, MPs are a subset of the Volterra series commonly used to lower the complexity of the models [17]- [23]. Also, memory polynomials have been used as part of multi-box structures to achieve high accuracy with lower model complexity than the standalone memory polynomial model [24], [25].
In this paper, new behavioral models dedicated for high efficiency Doherty power amplifiers (DPAs) are introduced. This is of prime importance since many of power amplification systems deployed in base stations are based on the DPA architecture. In fact, power efficiency of wireless communication systems is an essential part in the green communication systems. However, achieving high DPA efficiency introduces strong nonlinear effects due to the efficiency-linearity dilemma in PAs. Since linearity is a must for avoiding adjacent channels interferences, efficient amplification along with excellent linearization technique must be implemented. Many models in the literature deal with the PA modelling problem as a unified black-box model. Meaning, equations are used to determine the behavior of the PA along with its memory effects based on the measured input and output data. Those models range from different forms of complex-valued polynomial models to real-valued neural network-based models. The issue is that different PAs have different levels of memory effects. As such, when running an optimization algorithm, large number of fitting parameters can contribute to memory effects even though a PA's memory effect might be mild. In contrast, a large number of parameters related to memory effects would contribute to static characteristics of the PA. Optimization algorithms like genetic algorithm and particle swarm can only guarantee convergence to global minimum error without understanding the physical relevance of the parameters being optimized. Therefore, a certain number of parameters will exist in the model that does not affect the overall performance of the PA model while unnecessarily increasing the model complexity. The principal contributions of this paper are as follows: 1) Develop low complexity and high performance models for the memoryless AM-AM and AM-PM characteristics of GaN based DPA. 2) Integrate the static AM-AM and AM-PM models with memory effects functions to further enhance the modeled characteristics of the DPA. Two methods are proposed for integrating the static models along with the memory effects. The first method uses memory polynomials, and the second employs a deep neural network (DNN).
3) The proposed DPA models are more resilient, in the sense that their accuracies remain superior to the state of the art models following variations in the validation waveform. 4) When MP is used to model the memory effects of the device under test (DUT), the proposed model achieves more than 8dB enhancement in the normalized mean squared error (NMSE) at similar complexity when compared to state of the art MP models. 5) The use of DNN to model the dynamic distortions allows the proposed model to outperform standard NN models by approximately 3dB in the NMSE while having 85% to 90% less complexity.
The paper is organized as follows: first the proposed DPA models will be discussed. Then, the proposed modeling process is described and the obtained results are reported and discussed.
II. PROPOSED MODELING OF DPA
The nonlinearity at the compression point of the PA must be compensated for a device operating at reasonably high efficiency. A second nonlinearity exists in DPA that further worsen the linearity performance. Typically, a DPA is designed using Class AB and Class C amplifiers serving as the carrier and the peaking amplifiers, respectively. At relatively low input power, the Class AB amplifier becomes the dominant operator in the DPA. When input power reaches the compression point of the Class AB amplifier, Class C starts contributing to the output power of the DPA using load modulation. The second nonlinearity exists when both the class AB and class C PAs contribute to the output power.
The power characteristic of a DPA can be divided into several regions: first is Class AB linear region followed by Class AB compression region. Then, Class C turn-on region, and finally Class C compression occurs. This nonlinear behavior is depicted in Figure 1. Therefore, a model should predict Class C turn-on effect to comprehensively predict the behavior of the DPA.
A. STATIC DISTORTIONS MODELING
The static distortions sub-model is designed to have the ability to model the specific nonlinearity profile of DPA which commonly includes two inflection points as depicted in Figure 1. Furthermore, the proposed static model predicts the complex-valued gain behavior of the DPA by decoupling the AM-AM and AM-PM distortions. Equations (1) to (5) show the mathematical model for the complex-valued static gain, G PA , and the predicted signal, x out_SD , at the output of the static distortions model. (1) x out_SD,n = G PA,n G PA,n · x n (5) a 1 through a 7 , and b 1 through b 9 are real-valued fitting parameters, and x n is the complex-valued input sample at time instant n. One of the advantages of using the proposed model is that the number of parameters used in modeling the static characteristics of the DUT is always constant (16 parameters) regardless of the DPA used. This is true as long as the DPA exhibits the nonlinearity profile shown in Figure 1 which is standard for all high-efficiency DPA designs.
A careful examination of the AM-AM and AM-PM characteristics of Gallium Nitride (GaN) based Doherty power amplifiers brings to light a key observation: the inflection points of the AM-AM and AM-PM characteristics do not occur at the same input power levels. This behavioural misalignment in the distortions profiles of the AM-AM and AM-PM characteristics of GaN Doherty power amplifiers is not specific to the considered DUT but can be observed in several designs reported in the literature [25]- [28]. Therefore, it is anticipated that modelling the static AM-AM and AM-PM characteristics separately will enable a better independent control on the distortions profile and consequently lead to enhanced performance. In the proposed model, the independent modelling of the AM-AM and AM-PM characteristics is implemented using real-numbers based fitting parameters to model each specific characteristic of the DPA without affecting the other one. For example, fitting parameters in the proposed AM-AM model in Equation (1) do not affect AM-PM conversion characteristic. Also, fitting parameters in the proposed AM-PM conversion in Equations (2) and (3) do not affect the AM-AM characteristic of the DPA. Meaning, each characteristic can be separately modeled.
B. STATIC DISTORTIONS/MEMORY POLYNOMIALS MODEL
The proposed static distortions model is not suitable on its own, since modern PAs unavoidably exhibit memory effects. Hence, residual nonlinearities are expected between the measured output signal and the estimated signal at the output of the static distortion function. This residual memory effects signal is given by: where x out_meas,n is the measured output sample of the device under test.
In order to augment the proposed static model, a first configuration is proposed by combining the static distortions sub-model with a memory polynomials sub-model in order to accurately mimic the behavior of the Doherty amplifier. This augmented model is referred to as the SMP model (Static and Memory Polynomial model) since the static model output obtained from (5) is coupled with the memory polynomial output which augments the static model by including the memory effect as illustrated in Figure 2. The overall output of the SMP model is given by: The output samples x out_MP,n of the memory polynomial is calculated using: h jk are the complex-valued model parameters of the MP submodel. The complexity of the memory effects is controlled by the dimensions (nonlinearity order N MP and memory depth M MP ) chosen in Equation (8). It is important to note here that the MP function used in this model does not include the static term (hence k min = 1). This is to avoid redundancy since the static behavior is already modeled by the proposed static AM-AM and AM-PM functions. VOLUME 8, 2020 Figure 3. In this work, the DNN used has four layers: an input layer, two hidden layers and one output layer. The activation function used is a tanh function. Multiple activation functions other than tanh functions were tested, but they did not provide a usable model convergence. The input consists of 6 delays of the input x n , 6 neurons per hidden layer and two outputs for in-phase and quadrature components. The output of the static model x out_SD,n and the output of the neural network x out_DNN ,n are added together to yield the predicted output sample of the DPA. This model was optimized using the Keras API of TensorFlow platform for machine learning. The advantage for using a neural network rather than memory polynomial functions for modeling the memory effects will be discussed in the next section while assessing the relative performances of the two proposed models.
III. DISCUSSION AND RESULTS
In order to develop the proposed models, the static model should be separately identified, then the memory effects model will be derived. The identification process of the entire model is as follows: 1) Extract the memoryless AM-AM and AM-PM characteristics of the device under test by processing the raw measured data. This includes the steps commonly used in the identification of two-box models such as the augmented-Wiener, the Augmented-Hammerstein, and twin-nonlinear two-box models. These intermediate steps are namely time-delay alignment between the input and output waveform, data averaging to eliminate memory effects, and finally calculation of the memoryless gain magnitude and phase by using the averaged time-aligned versions of the measured input and output baseband complex waveforms. 2) Apply fitting algorithm to calculate the parameters of the static AM-AM model (a 1 through a 7 ) from equation (1), 3) Apply fitting algorithm to calculate the parameters of the static AM-PM model (b 1 through b 9 ) from equations (2) and (3). 4) Use Equations (5) and (6) to de-embed the measurement data and obtain the desired complex output signal of the memory effects modelling function x out_ME . 5) Identify the parameters of the second sub-model using the input signal and the desired sub-model output signal x out_ME . This step differs whether the model being identified is the SMP or the SDNN.
The flowchart in Figure 4 summarizes the parameters identification process described above. As it is the case for any behavioral model, the derived coefficients and the model accuracy is only valid for the observed behavior. Variations in the test signal or operating conditions which will emulate a different behavior would require a new identification of the model coefficients as per the steps described above. This is common to any behavioral model and does not affect the generality of the proposed model.
For the experimental validation of the proposed model, a Doherty PA (DPA) employing a 10W GaN transistor is used as the device under test. In this DUT, the carrier amplifier was optimized for efficiency using harmonic tuning. The PA operates in the 2425MHz frequency range. The DUT was driven by a 20MHz long term evolution (LTE) test signal having a peak to average power ration (PAPR) of 9.7dB. The instantaneous input and output baseband complex waveforms were acquired using the standard setup for power amplifiers characterization using modulated test signals [5] in which an arbitrary waveform generator was used to generate the RF test signal applied at the input of the DUT. The output signal was attenuated and then digitized using a commercial vector signal analyzer.
The measured AM-AM and AM-PM characteristics of the DUT as well as their estimated static versions are reported −15dBm for the DUT), the gain exponentially decreases due to saturation of the DPA. Therefore, a sigmoid function is used to model the saturation of the DPA. Fitting parameters a 5 and a 6 represent the level and the rate of the gain decay, respectively. Fitting parameter a 7 represents the input power at which the gain starts to exponentially decrease. • In the middle input power region (−45 to −20dBm for the DUT), the gain decreases (from −45 to −25dBm for the DUT) and then increases again (from −25 to −20dBm for the DUT) due to the compression of the main amplifier followed by the turn-on of the peaking amplifier. This is modelled using a bell-shaped function to mimic the decrease and increase of the gain at this region. Fitting parameter a 2 represents the amount of gain decrease in this transition region while fitting parameter a 3 represents the input power at which the peaking amplifier starts contributing to the DPA output power. The reasons behind the proposed formulation of the static AM-PM model are as follows: • Fitting parameter −b 4 represents the constant phase distortion of the carrier amplifier of the DPA at low input power (−50 to −40dBm for the DUT).
• Then phase distortion starts to increase due to gain compression of the carrier amplifier (−40 to −23dBm for the DUT). This modelled using an exponential function. Fitting parameter b 5 represents the level of the increase of the phase distortion in the carrier amplifier in this VOLUME 8, 2020 power region. Fitting parameter b 6 represent the rate of rise of phase distortion.
• At higher input power (more than −25dBm for the DUT), the peaking amplifier kicks in, and the effect of the carrier amplifier starts to diminish. Therefore, a function is needed to start suppressing the phase distortion of the carrier amplifier. In this work, transition function as formulated in Equation (3) is used to do so. This function, f i , has three fitting parameters: b 1 which represents how fast this transition functions saturates, b 2 represents the level of saturation, and b 3 is the knee point for the transition function.
• To model the effect of the peaking amplifier compression (from −23 to −15dBm for the DUT), a second exponential function is used to model the phase distortion. This part has three fitting parameters: b 7 represents the level of the increase of the phase distortion in this power region, b 8 represent the rate of rise of the phase distortion, and b 9 represents the input power at which the peaking amplifier kicks in.
The AM-AM and AM-PM models' build-up is depicted in Figure 6. This figure clearly illustrates the effect of each term in Equations (1) In order to benchmark the performance of the proposed model against previously reported state of the art model, the NMSE was used. The NMSE is commonly used to benchmark the performance of behavioral models, and as is defined as: x out_meas,i 2 (9) where L refers to the number of samples used for calculating the NMSE. x out_meas,i and x out_est,i are the i th measured and estimated output samples of the validation waveforms, respectively.
The NMSE was calculated using 40,000 samples. The training NMSE refers to the NMSE obtained when the validation waveform is the same as the waveform used to train the model.
The SMP model was benchmarked against several comparable models including the standalone MP model, the EMP model [21], and a hybrid model made of the combination of the MP and EMP models [29]. The training NMSE results reported in Table 1 clearly demonstrate the effectiveness and superiority of the proposed model as it achieves significantly lower NMSE than the benchmark models while using comparable number of parameters. Indeed, the proposed SMP model reduces the NMSE by approximately 8 to 12dB as reported in Table 1.
The SDNN model was also compared to other state of the art neural network based models. The parameters of the DNN used in the SDNN are the same as those described in section II. In table 2, the training NMSE of the proposed SDNN model is compared to that of the NN based models reported in [9], [24] and [25]. The results reported in this Table reveal the ability of the proposed SDNN model in achieving lower NMSE while requiring significantly less parameters resulting in between 85% and 90% less complexity without compromising the model accuracy. To further assess the accuracy of the SDNN model and its ability to accurately predicted the amplifier's output signal, the measured spectrum at the output of the DUT as well as the estimated spectrum at the output of the SDNN model are reported in Figure 9. These plots corroborate the accuracy of the SDNN model as expected by its NMSE performance. Moreover, the adjacent channel leakage ratio (ACLR) was calculated at the output of the DUT and its SDNN model. These results reported in Table 3 confirm the accuracy of the model and its ability to precisely predict the behavior of the DUT in time domain (as confirmed by the NMSE results) and frequency domain (as illustrated by spectra and ACLR data).
Furthermore, to evaluate the robustness of the considered models, 10 other waveforms (different from the training waveform) were used to validate the models. The original test signal contained 200,000 samples, 40,000 of which were used to train the models. The 10 validation waveforms were selected as part of the remaining 160,000 samples of the test waveforms that were not used to train the model. Hence, a total of 11 NMSE values (1 training NMSE and 10 validation NMSE) were calculated for each model. The results of this robustness assessment are summarized in Figure 10. This figure reports the training NMSE for each model, as well as VOLUME 8, 2020 the maximum, minimum and mean validation NMSE of each model based on the results of the 10 validation waveform. For consistency of the results, the same 10 validation waveforms were applied to all models. Based on this data, it appears that the proposed SMP and SDNN models outperform all other models. Moreover, the SDNN models has a better robustness than the SMP model as it can be observed by the reduced NMSE variation following changes in the validation signals. However, this is achieved at the expense of a higher model complexity.
IV. CONCLUSION
In this paper, novel behavioral models dedicated for GaN based Doherty power amplifiers were proposed. The novelty resides in the use of novel polar domain independent modeling of the static distortions of the DUT. For this purpose, independent functions were used to model the memoryless AM-AM and AM-PM characteristics of the device under test. This enabled a more accurate modeling of the two distortions profiles which showed inflection points occurring at different input power levels. The proposed static distortions model was then augmented with a memory polynomials, and later by a DNN function. Experimental validation demonstrated that the proposed models achieved better NMSE than comparable state of the art models. The proposed SMP model achieved between 8 and 10dBs better NMSE than its previously reported counterparts. Furthermore, the SDNN model led to slight NMSE enhancement compared to other NN based models but up to 90% reduction in the model complexity. Finally, a study of the robustness of the proposed models was carried out and showed that the performance of the SDNN model is less sensitive to variations in the validation signal than that of the SMP model. | 5,528.6 | 2020-01-01T00:00:00.000 | [
"Computer Science"
] |
Y‐Stabilized ZrO2 as a Promising Wafer Material for the Epitaxial Growth of Transition Metal Dichalcogenides
Y‐stabilized ZrO2 (YSZ) as a promising single‐crystal wafer material for the epitaxial growth of transition metal dichalcogenides applicable for both physical (PVD) and chemical vapor deposition (CVD) processes is used. MoS2 layers grown on YSZ (111) exhibit sixfold symmetry and in‐plane epitaxial relationship with the wafer of ( 101¯0$2 \bar{2} 0$ ) MoS2 || ( 2¯11$\bar{2} 11$ ) YSZ. The PVD‐grown submonolayer thin films show nucleation of MoS2 islands with a lateral size of up to 100 nm and a preferential alignment along the substrate step edges. The layers exhibit a strong photoluminescence yield as expected for the 2H‐phase of MoS2 in a single monolayer limit. The CVD‐grown samples are composed of triangular islands of several micrometers in size in the presence of antiparallel domains. The results represent a promising route toward fabrication of wafer‐scale single‐crystalline transition metal dichalcogenide layers with a tunable layer thickness on commercially available wafers.
Introduction
Transition metal dichalcogenides (TMDCs) are considered as promising next-generation optoelectronic materials, owing to their intriguing physical properties appearing as a result of an intrinsically layered structure, specific crystal symmetry, and distinctly different optoelectronic properties of a large variety of polytypes. [1,2]To fully understand and explore the potential of TMDCs, single-crystalline wafer-scale films with tuneable material composition, layer thickness, and crystalline excellence comparable to, or even exceeding that of currently available bulk crystals, are needed.However, synthesis of the films, satisfying these requirements, remains very challenging.Despite van der Waals nature of the materials, growing them on available non-van der Waals wafers leads to a strong nuclei-wafer interaction, predefining structural imperfection of the deposited layers.The choice of suitable substrates for the epitaxial growth of TMDCs is limited not only by purely crystallographic considerations, but also wafer resistance against melting, decomposition, and chemical reactions with chalcogen atoms at elevated temperatures.These conditions are required for the most widely used fabrication techniques, including physical vapor deposition (PVD) and chemical vapor deposition (CVD).The commonly used basal plane of sapphire is thermally and chemically stable and withstands even the harsh environment of the CVD process.However, the in-plane lattice mismatch between classical TMDCs (e.g., 2H-polytype MoS 2 a MoS2 = 3.161 Å, c MoS2 = 12.295 Å [3,4] ) and sapphire (a Al2O3 = 4.76 Å, c Al2O3 = 13.00Å) is huge (about À34%).As a result of coincidental matching of several lattice units, two types of in-plane rotational domains induced by domain-match epitaxy have been reported.The first configuration corresponds to the 3  a TMDC % 2  a Al2O3 supercell matching.The second variant involves a 30°(or 90°) rotation of the TMDC lattice with respect to sapphire and larger supercell consisting of five TMDC units (5  a TMDC % 2 ffiffi ffi 3 p  a Al2O3 , R30°configuration). [5][8] Despite the fact that the excellent device performance for the layers grown on sapphire has been successfully demonstrated, [9] there are very few reports on single-crystalline wafer-scale growth that have appeared only recently. [5,10,11]The large mismatch between MoS 2 and sapphire becomes even more severe for layers grown by PVD, where the growth temperatures are several hundred degrees lower, compared to the CVD technique.In this case, the surface diffusion of adatoms is low, and the typical domain size is in the tens-nm range. [12]Merging the individual crystallites in a quasicontinuous film without preserving long-range epitaxial registry leads to a high density of grain boundaries, negatively affecting optical and electrical properties of the layers.As a possible alternative to sapphire, we utilize cubic YSZ (111) as commonly available substrate for TMDC growth.Here, we demonstrate the suitability of YSZ for epitaxial growth of MoS 2 with pure sixfold symmetry for both PVD and CVD processes.DOI: 10.1002/pssr.202300141Y-stabilized ZrO 2 (YSZ) as a promising single-crystal wafer material for the epitaxial growth of transition metal dichalcogenides applicable for both physical (PVD) and chemical vapor deposition (CVD) processes is used.MoS 2 layers grown on YSZ (111) exhibit sixfold symmetry and in-plane epitaxial relationship with the wafer of (1010) MoS 2 || (211) YSZ.The PVD-grown submonolayer thin films show nucleation of MoS 2 islands with a lateral size of up to 100 nm and a preferential alignment along the substrate step edges.The layers exhibit a strong photoluminescence yield as expected for the 2H-phase of MoS 2 in a single monolayer limit.The CVD-grown samples are composed of triangular islands of several micrometers in size in the presence of antiparallel domains.The results represent a promising route toward fabrication of wafer-scale single-crystalline transition metal dichalcogenide layers with a tunable layer thickness on commercially available wafers.
Results and Discussion
ZrO 2 can be stabilized by Y as a face-centered-cubic fluorite structure with a lattice parameter depending on the Y-molar fraction.The material is generally known as a substrate for In 2 O 3 -based bixbyite compounds. [13]A closely packed YSZ (111) cubic substrate (a = 5.147 Å, Crystec GmbH) with a nominal Y doping level of 9 mol%) exhibits an in-plane lattice spacing along the [110] direction of a YSZ 110 = 3.639 Å.This coincides with the in-plane lattice parameter a MoS2 of 2H-MoS 2 with a lattice mismatch of about -13.1%.At first, we address the PVD growth of multilayer and single-monolayer (ML) films.
The multilayer samples were structurally characterized using high-resolution X-ray diffraction.The corresponding out-ofplane measurements (2θ-ω scan) show a prominent peak at about 2θ = 14.0°, which can be assigned to the MoS 2 00.2 Bragg reflection (Figure 1a).The resulting vertical lattice parameter of c exp = 12.7 Å is in approximate agreement with the literature value of c MoS2 = 12.3 Å. [4] In the immediate vicinity of the Bragg reflection, pronounced thickness fringes are observed, indicating a rather smooth interface between the substrate and the layer.From the angular spacing of the fringes, we can determine a MoS 2 film thickness of t = 34.8(7)Å, which corresponds to about five MLs.A rocking curve (ω-scan) carried out at 2θ = 14.0°(Figure 1b) shows a full width at half maximum value of about Δω = 51 arcsec, which is close to the experimental resolution of the instrument.This shows that the MoS 2 layer was planar on the YSZ substrate.In-plane measurements under grazing incidence (GIXD) were performed to determine the full epitaxial relationship between the MoS 2 layer and the YSZ substrate.Figure 1c shows two in-plane rocking curves (ϕ-scans) taken at 2θ = 32.9°(corresponding to the MoS 2 10.0 Bragg reflection) and at 2θ = 49.7°(corresponding to the YSZ 220 Bragg reflection).From these curves, it is first clear that the layer and substrate exhibit sixfold in-plane symmetry.The width Δϕ sub = 0.1°of the substrate reflections matches approximately the experimental angular resolution, while the layer reflections are noticeably wider (Δϕ layer = 1.8°).However, this in-plane mosaic spread is more than a factor of 2 smaller than for MoS 2 layers grown on sapphire (Δϕ layer = 4.2°), see Figure S1, Supporting Information.Besides the observed strong layer reflections, no other Bragg reflections attributable to the MoS 2 layer appear, that is, no 90°variant is observed.This underlines a clear epitaxial relation between layer and substrate.From the relative position of layer and substrate reflections, it is evident that the MoS 2 1010 net-planes are oriented parallel to the YSZ (211) net-planes.This epitaxial relationship is outlined in Figure 1d.The in-plane lattice parameter a exp = 3.14 Å derived from 2θ/ϕ-scans, is very close to the literature value for bulk MoS 2 , indicating that the layer is fully relaxed.Atomic force microscopy (AFM) images of the sample are shown Figure S2, Supporting Information.
Decreasing the layer thickness down to 1 ML is expected to result in the appearance of a photoluminescence (PL) band, corresponding to the transition from the indirect to direct band in the 2H-polytype of MoS 2 . [14] AFM image (Figure 2b) reveals MoS 2 islands with a lateral size of up to 100 nm.The islands are aligned along the step edges of the substrate, which have a miscut of about 0.4°toward the [132] direction.The height of the islands of %0.7 nm is consistent with the thickness of a MoS 2 monolayer.The corresponding PL spectrum (Figure 2c) is composed of the PL band of MoS 2 (centered at 664 nm) with contributions of the YSZ wafer at 550 nm and 820 nm.It has to be noted that the tiny size of the crystallite domains, structural imperfection, and phase mixing during initial stages of the growth are the typical problems of PVD of TMDCs on non-van der Waals wafers. [15,16]For this reason, the well-ordered structure of MoS 2 nuclei, together with observation of the bright PL signal from the submonolayer thin film, represents a promising result, stimulating further process optimization.
The chemical and thermal stability of YSZ in the form of a thin and flexible ceramic substrate planarized with SiO 2 was recently demonstrated for CVD growth of few-layered MoS 2 films using ammonium heptamolybdate and sodium hydroxide as precursors. [17]In order to further prove the suitability of YSZ as a single-crystalline substrate for epitaxial growth of TMDCs, we performed a number of dedicated CVD growth experiments which are summarized in Figure 3.The layer orientation verified by RHEED coincides with that found in the PVD samples (Figure 3a).Optical microscopy reveals the formation of nonideal triangular MoS 2 domains with a size of 5-10 μm (Figure 3b).The single ML thickness and the 2H-phase of the domains are confirmed by PL measurements and AFM imaging.The PL spectrum of the sample shown in Figure 3c represents again a superposition of the signal from the layer and the substrate (red curve), while the measurements on the area in between the islands (black curve) depict the response of the substrate only.Optical microscopy clearly shows preferential in-plane epitaxial orientation of the islands.However, it is worth noting that antiparallel domains-that is, 180°domains-also show up.A closer inspection of the sample by AFM reveals that one of the zigzag edges of the MoS 2 triangular is aligned perpendicular to the atomic steps of the wafer, appearing as a result of a substrate miscut of 0.2°toward the [110] direction (Figure 3d).Statistical analysis of optical microscopy images (Figure S4, Supporting information) reveals that this alignment holds for 93% of the domains.The second observed variant corresponds to a 30°(or 90°) rotation of the MoS 2 crystals.The visual rotation of the crystallites can be attributed either to a transition from zigzag to armchain edges in about 7% of the domains or to a 90°r otation of MoS 2 lattice with respect to YSZ.In the latter case, the epitaxial supercell is formed by four MoS 2 units which commensurate with the substrate as 4  a MoS2 %2 ffiffi ffi 3 p  a YSZ 110 with a lattice mismatch of about 0.3%.Despite the significantly smaller lattice mismatch compared to the in-plane alignment of MoS 2 (1010) || YSZ (211), the R30°variant appears to be energetically unfavorable, as it was rarely observed in the CVD samples and not detected in the PVD-grown films.
The results on PVD-and CVD-grown samples confirm that the in-plane epitaxial relationship between the MoS 2 layer and the YSZ substrate is not affected by the small random miscut direction of the wafer that results from wafer polishing.However, when depositing a material with threefold symmetry on a substrate with sixfold symmetry, it is expected from simple crystallographic considerations that 180°domains will be present in the layers.This can be clearly seen in the CVD-grown films.Note that both domain orientations are found with virtually equal incidence; however, these cannot be distinguished in our GIXD measurements.The identical scenario was recently reported for MoS 2 and WS 2 thin films deposited on c-plane sapphire.In that case, the degeneracy in the nucleation energies for parallel and antiparallel domains was successfully lifted by selecting a miscut orientation where the substrate step edges are aligned parallel to the zigzag edge of the triangular TMDC islands. [5,10]he observed similarity between sapphire and YSZ in terms of domain nucleation and growth process provides excellent perquisites for the fabrication of single-oriented TMDC layers on the substrate with smaller lattice mismatch.Careful design of the wafer miscut orientation in combination with appropriate growth stoichiometry plays an important role in this process.
Conclusion
We introduce Y-stabilized ZrO 2 (111) as a commonly available wafer for the epitaxial growth of TMDCs.The material has been found to be applicable for both PVD and CVD growth processes.MoS 2 layers grown on YSZ (111) with a small miscut toward the [132] and [110] directions exhibit sixfold symmetry and an inplane epitaxial relationship with the wafer of (1010) MoS 2 || (211) YSZ according to GIXD and RHEED.Optical microscopy of the noncontinuous CVD layer reveals the presence of 180°r otation domains, as expected from epitaxial growth of a threefold material on a sixfold substrate.The observed similarity of the domain formation between sapphire and YSZ offers great opportunities for the fabrication of single-oriented TMDC layers on the substrate with smaller lattice mismatch by careful design of the wafer miscut orientation, combined with appropriate growth stoichiometry.The results are especially promising for the PVD process, where nucleation of MoS 2 islands with a lateral size of up to 100 nm and strong PL as expected for 2H-phase of MoS 2 in a single monolayer limit is observed already during the first growth runs.
Experimental Section
The epitaxial growth was performed on 10 Â 10 Â 0.5 mm 3 YSZ (111) wafers (Crystec GmbH).The wafers were cleaned with organic solvents and annealed under air at 1000 °C for 2 h prior to the growth.In the case of PVD samples, a 500 nm-thick Mo layer was deposited on the back side of the substrate in order to improve the heat absorption and enable homogeneous heating of the sample.The growth temperature (T G ) was controlled by an infrared pyrometer with an emissivity factor of 0.7.The PVD samples were grown using pulsed thermal deposition technique. [18] Mo wire with a diameter of 1.25 mm (GoodFellow, purity 99.9%) was used as a filament.Sulfur (S) was supplied by a valved cracker source (MBE Komponenten GmbH).The growth was performed via sublimating Mo in a pulsed mode under continuous exposure of the substrate to sulfur.The pulse duration and Mo sublimation period were set to 90 and 180 s, correspondingly.The sublimation current was adjusted to provide nominally single-ML equivalent Mo coverage during five pulses.The YSZ wafers were preheated prior to the growth under the S flux at 850 °C for 10 min.The deposition was performed at a growth temperature of T G = 620 °C and S equivalent beam pressure of 3 Â 10 À7 Torr.The CVD sample was grown in a horizontal tube furnace with Ar carrier gas.MoO 3 powder was used as precursor.The CVD temperature was T CVD = 850 °C.Raman and PL measurements were performed with a Horiba confocal Raman setup using a 532 nm laser line for the excitation and 1800 mm À1 optical grating.AFM was conducted with a Bruker Dimension in a ScanAsyst mode.The high-resolution X-ray diffraction experiments were performed on a 9 kW Rigaku SmartLab diffractometer.For the out-of-plane measurements, a parabolic multilayer mirror was used in combination with an asymmetric Ge 220 channel-cut crystal providing a monochromatic (λ = 1.54056Å) and collimated (29 arcsec) X-ray beam.A Hypix 3000 2D area detector was used in 0D mode as a point detector.For the GIXD in-plane X-ray diffraction experiments, Soller slits were used for the in-plane collimation of the incident (0.15°) and diffracted (0.228°) beams.Typical glancing angles of incidence on the samples were below the critical angle of total external reflection to ensure maximum sensitivity to the MoS 2 film, typical value being 0.2°.
Figure 1 .
Figure 1.Structural properties of a 3.5 nm MoS 2 layer grown by PVD on YSZ (111) wafer.a) Out-of-plane X-ray diffraction 2θ-ω scan in the vicinity of the MoS 2 00.2 Bragg reflection.b) Corresponding experimental X-ray rocking curve (ω scan) at the MoS 2 00.2 Bragg reflection fit with a Gaussian line profile (red line).c) GIXD in-plane X-ray diffraction rocking curves (ϕ-scans) of the YSZ wafer (blue) and the MoS 2 layer (red), taken at 2θ = 32.9°(MoS 2 10.0 Bragg reflection) and at 2θ = 49.7°(YSZ220 Bragg reflection).d) Epitaxial relationship scheme of YSZ wafer (marked in blue color) and MoS 2 layer (marked in red color).
Figure 2a displays reflection highenergy electron diffraction (RHEED) patterns from the bare YSZ wafer (top) and after deposition of about 0.5 ML-thick MoS 2 equivalent (bottom).The RHEED patterns show only one system of the streaky reflections from MoS 2 (shown by the blue lines with corresponding Miller indices in Figure 2a), indicating a 2D growth mode and a clear epitaxial relationship with the substrate (reflections of the YSZ substrate are highlighted by the red arrows) with sixfold symmetry, as previously verified for the multilayer sample.The relative distance between the MoS 2 streaks (blue lines) and the YSZ streaks (red lines) indicates that the MoS 2 layer is relaxed already in the first monolayer.
Figure 2 .
Figure 2. Submonolayer thin MoS 2 film grown by PVD on (111) YSZ.a) RHEED patterns of bare YSZ substrate (top, marked by blue lines) and MoS 2 layer deposited on YSZ (bottom, red lines).The dashed lines correspond to the surface reconstruction.The RHEED patterns were recorded along the [110] and [211] directions of the YSZ wafer.The Miller indices of the MoS 2 Bragg reflections (highlighted by red lines) are labeled.b) AFM image of the sample.The inset shows the in-plane orientation of the wafer and height profile along the black line.c) PL spectrum of the sample excited with a 532 nm laser line.The PL/Raman bands at 550 nm and 820 nm originate from the YSZ wafer.
Figure 3 .
Figure 3. MoS 2 layer grown by CVD on YSZ (111) substrate.a) RHEED patterns of the sample recorded along the [110] and [211] directions of the YSZ wafer.Bragg reflections of the YSZ wafer and MoS 2 are highlighted by the red and blue lines, correspondingly.b) Optical micrograph, demonstrating in-plane preferential orientation of 1 ML-thick MoS 2 domains.c) PL spectra, from a MoS 2 island (red curve) and from the bare substrate surface (black curve).d) Overview AFM image of the triangular domain and the height profile along the black line.The blue line depicts the direction of atomic step edges of the wafer. | 4,242 | 2023-06-25T00:00:00.000 | [
"Materials Science"
] |
China’s FDI Impact on Economic Development in the Developing Countries in Africa (The Case Study of Comoros)
China's continued development assistance to Africa in recent decades is one of the most controversial events in the region. China's aid to Africa, fueled by the spirit of the growing self-secession of African countries, has attracted the world's attention. While friends see the great help as a lucrative opportunity for both sides, the enemies of the association fear the potential impact of Chinese development aid on increasing their own stoic capacity. Despite the growing debate between friends and enemies of these associations, little existing literature has examined the sincerity of China's capacity-building assistance, focusing on job creation and technology or ICT transfer. There is a lack of clear evidence of capacity building indicators, such as job creation and technology transfer. In the absence of literature, this study tries to emphasize and examine China's impact on the construction of chamber capabilities in job creation and technology transfer. The study used key data from semi-structured interviews that included open questions and secondary data, such as the White Paper in China. Unlike previous studies in Angola, Sudan and Uganda, the situation in the Comoros is quite the opposite. The results paint an image of the impact of support for the Chinese project on the construction of capacities in the Comoros, which at best remains "negligible." The country can reach its full potential and there is an urgent need to restore the political and economic power of Comoros.
Introduction
This excerpt from China's Foreign Ministry spokesman, Jiang Yu's response to a joint question-and-answer session following recent criticism of China's foreign policy and commitment in Africa, in which he described China's assistance in Africa as a trust booster, captures the main reason for this investigation. This issue has also caught up with my interest in launching controversial debates and discussions in academic, media and political circles on the nature of development aid by China, to African countries, which are the main beneficiaries, the cause of this aid and the impact of this aid on the beneficiary countries. This has finally led to the abundance of literature with polarized ideologies in this area of controversy. While Taylor (2004) agrees that the main impetus behind China's interest in Africa is economic and diplomacy, Shelton (2001) represents China's commitment to Africa, driven by the three main strategic interests; Africa offering ready-to-market Chinese products, resource gap, and oil that beautifies China's rise. China plans to hide the imperialist formulation toward the African counterpart in order to embed all dividends on the way, according to Rupp (2008).
On the other hand, some scientists, as well as officials from most recipient countries and Chinese government officials, are destroying this China -Africa relationship. Alternatively, viewing China's assistance to Africa from the point of view of interdependence, there are claims that China's development aid offers a lucrative opportunity for both sides. Although previous investigations into cases such as Sudan, Angola and Uganda have spurred some boost in capacity building because they are committed to China, the Comoros case was different. The results of the Comoros case show the impact of Chinese project support on capacity building in the Comoros, which is at best "negligible". Unlike the cases in Sudan and Uganda, which have some impetus to develop capacities because of their relations with China, the same could not have been done to help the Comoros.
This conditionality as well as the general formalities such as the principle of non-interference that surrounds China's aid to these recipient countries engulfs the part of the problem addressed in this research. One of such conditionality that has greatly attracted my interest and it is my responsibility to research is the fact that most of the implementation and constructions of these projects are mostly carried out by Chinese companies. With the implementation process of most of these aid projects effectuated by Chinese companies, it has developed a huge debate for critics, over the continuous utilization of Chinese labour and input by these companies, which undermines the economic prospect of its relation with the recipient country. With such undermining prospect, critics such as renowned African economist George Ayittey, have questioned the rhetoric of China's aid being a capacity builder when the nature of these aids deal are not transparent, their impact on local economies is outrageous and they offer scanty employment opportunities to the recipient countries such as the case of Uganda. With that in mind, this research will seek an in-depth case analysis of impact of this China's development aid towards capacity building of Comoros, while hoping that the findings from this research will assist academicians and politicians to design policies to enhance Sino -African relationship.
Research Questions
Fascinated by the perspectives of actors who use a realistic approach to describing China, Africa's partnership undermines Africa's capacity-building in the Chinese principle without interference in issues that significantly undermine human rights, good governance, democracy and the continued use of Chinese workers and have not documented any information detailing their aid relationships, I am very excited to explore the empirical basis of the impact of these subsidies on economic development. In this respect, I anticipated this research question to be answered: How does Chinese aid contribute to capacity building in Comoros in job creation and Technology or ICT transfer?
Significances of Study
The research is a bit of an attempt to add to the ongoing conversation and debate in academia on various issues surrounding China's support for Africa. The document will examine China's assistance in one country and contribute to further literature that can be used, unlike other host countries. Moreover, I hope that by encouraging quest and interest for this research we can finally recruit more brains to ask really relevant and interesting questions about the current situation surrounding China's honeymoon with Africa.
Comoros Economic Development Overview
The Comoros archipelago is located in the Indian Ocean, north of the Mozambique Channel and northeast of Madagascar. While the spread of the virus in the Comoros so far appears to have been contained, economic activity has slowed sharply prior to the COVID-19 pandemic (coronavirus).
The current submissions of the EUP in the Comoros remain marginal, despite tax breaks and benefits offered by the Government, which aims to resolutely improve the climate and institutional stability of the country. According to UNCTAD's World Investment Report 2020, FED bookings in 2019 were only $8 million, which is in line with historical trends, but well below the exceptional peak in 2011 (EUR 23 million). AED shares remain low despite an increase in 2019, estimated at $129 million. The agricultural and fisheries sectors receive the largest share of foreign investment. The main investment countries are France, the United States and South Africa.
Comoros ranked 160 of the 190 countries in the World Bank's Doing Business Report 2020, gaining four places compared to the previous year. In May 2018, the African Development Bank approved a payment of USD 1.5 million for a support project from an investment promotion agency (PAAPI). Comoros' investment code does not provide an explicit definition of EDI and the regulatory environment continues to impose significant burdens on investors. However, the government is determined to reform the investment code and establish a single point of contact for the creation of a company. Poor infrastructure quality, limited internal market size, geographical isolation, frequent water and electricity shortages, limited natural resources and unseeded labour are also hampering EDI. In addition, property rights are not well protected and contracts are applied weakly. The legal system, based on both Sharia (Islamic) law and the French legal code, is weak and under political influence. Corruption is reported at all levels of government and exacerbated by internal political disputes and competition for resources between the administrations of the three islands. China is the country's largest investor. Over the past European Journal of Business and Management www.iiste.org ISSN 2222-1905(Paper) ISSN 2222-2839(Online) Vol.12, No.27, 2020 decade, the Chinese have invested millions in the country. They paved roads, built new schools, mosques, markets, government buildings (including the parliament building), a new airport, a centre to facilitate tourism and a sports stadium. In addition, China regularly sends doctors to the islands and builds a 12-million-dollar hospital. Table 1: Comparing protection of investors according to countries Note: *The higher the index, the more transparent the terms of the transactions. **The higher the index, the more the manager is personally responsible. ***The higher the index, the easier it will be for shareholders to take legal action. ****The higher the index, the higher the level of protection for investors.
China Foreign Aid to Africa
China's official development financing in Africa consists mainly of interest-free grants and loans, debt relief and soft loans, as well as preferential export loans, commercial loans and loans from Chinese banks. Subsidies (rarely granted as budget aid) and interest-free loans were the main instruments of the ODA granted by China until 1995, when soft loans were introduced (Brautigam, 2011). China oversees aid and delivery of most of China's external finance. Both banks act as part of the Beijing Toolkit to support China's development goals (Brautigam, 2011). Despite this criticism, researchers like Davies et al. (2008) acknowledge the controversial role Chinese aid plays in Africa, but they portray China's role in economic development in recipient countries like Angola and Sudan. They described China's continued relationship with both Angola and Sudan as mutual benefit. The government's policy of reducing the number of people living in poverty is a major challenge.
While Guloba et al. (2010) research on China's assistance to Uganda further cradles some of the great positive effects of Chinese aid to Uganda, they however went on to highlight some major concerns about China's assistance to Uganda, which includes marginal employment of locals and the continued exploitation of Chinese foreigners, where these projects are being implemented. Nevertheless, their results show that China's assistance, which is mainly in the form of technical assistance through training in Chinese institutions, grants, interest-free loans, preferential loans, debt relief, have all gone a long way to reducing Uganda's external burden and thus has had a positive impact on the well-being of the people of Uganda. Wang et. al (2010) further delineates such positive effects of Assistance to China by referring to significant funding for multi-purpose development projects in sectors such as electricity (hydropower) and transport (railways and motorways), which are sectors in beneficiary countries that attract relatively little help from traditional donors. However, very few studies in Comoros literature have focused on the impact of Chinese FDI to the Comoros.
Conceptual Analysis
For nearly five decades, several wealthy donors have provided aid to African countries. The motive for such assistance ranges from various development, economic, commercial and political perspectives that seek mutual benefit for both donors and recipients. However, these aid measures normally cover all financial transaction agreements between one government or actors to another.
It is undeniable that, on the basis of the above definition, aid links will include the transfer of funds, the supply of goods, projects, programmes, construction and debt relief to the recipient countries. In fact, foreign aid has become a focus and locus in the recipient countries (Aluko & Arowolo, 2010). This view eventually led to donor countries eventually being confirmed as a foreign policy equipment used to support and spread their influence over the beneficiary countries, which gave their role as the center to strengthen the definition of foreign aid to the OECD. Vol.12, No.27, 2020 This mindset is largely confirmed with the ideologies of Bretton Wood's institution. But with China becoming visible as a viable aid donor in Africa, the aid paradigm seems to be undergoing a make-over.
Theoretical Framework
A plethora of approaches and theories has thrived within the academic and political parameters to explain aid effectiveness. Ranging from Classical Realism, Dependency theory, Constructivism, Geopolitics theory, Globalism, Idealism and Imperialism, just to name few, all have sought to achieve a scientific explanation of the controversies surrounding aid effectiveness. Nonetheless, this research will adopt the Interdependence theory and the Political Realism which are two opposing theoretical views, predominantly employed by actors to explain the nature of the relation between China and Africa.
Interdependence theory is a broad and complex theory that was postulated during the 1970's, when the Political Realist perspectives on international relations was failing to take into account many of the new aspects of interstate relations (Brenner, 2000). Though the underlining settings of this theory have changed over time, the basic principles of this theory have remained the same. Interdependence is generally defined as mutual dependence between two entities. Created through the expansion of international transactions, with the benefits exceeding the cost, this description of Interdependence by Robert Keohane & Joseph Nye greatly captures the current nature and scope of the China -Africa relation in accordance with the arguments postulated by friends of this relationship.
According to friends of this relationship, China -Africa relation is a reciprocal relationship where both actors are dependent on each other. Though the scope of the relationship is not similar i.e. China seeking raw material and market for its product and Africa seeking development assistance, both actors are simultaneously dependent on each other. Both actors can affect each other through their own resources i.e. China has the financial resources needed by Africa while Africa on the other had has the natural resources and the market needed by China. This perception greatly captures (Enuka 2011, p 44, 53) depiction of the China -Africa relations. He aptly portrays this relation as Interdependent whereby he depicts Africa as a poor and underdeveloped actor in search for financial resources and China as a relatively better off actor who is in dire need of raw materials, market and support. Thereby, implying that if China provides the financial and infrastructure resources needed by Africa, the spillover effect with eventually lead to increased employment creation and technology transfer which are all capacity building indicators, that depicts growth.
However, it is important to be mindful of the fact that, this reciprocal relationship must not necessarily be symmetrical. Though depicted as mutually dependent on each other, one should be careful not to assume that Interdependence is an evenly balanced relationship. Robert Keohane & Joseph Nye (1977) used a very comprehensible example whereby actor A maybe dependent on actor B with respect to oil and actor B maybe dependent on actor A with respect to food. Though both actors are depending on each other, it is wrong to describe such a situation in which the control of actor A over actor B with respect to food is "balanced" by actor B control of actor A with respect to oil.
Formulated in accordance to DAC member's personal principles and notion of socioeconomic development, aids issued out are intended to orient the recipient countries in the direction preferred by the donors. This ideology eventually prompted Hans Morgenthau (1962: p 309) to argue that "policy of foreign aid is no different from diplomatic or military policy or propaganda; they are all weapons in the political armory of the nation". In this respect, the motive of aid does seem to be mix political and economic desires of donors (Greber, 2012). This eventually implies that, despite their provision of aids, they actually do not lay emphasis on the impact towards capacity building in these recipient countries, but seek to promote their national interest.
Research Design
It is doubtless that the research, with the above short description, will use a case study approach. As Yin (2003) aptly puts it, one should consider a case study design when the focus of the study is to answer how, when it comes to real-life behavior, that cannot be manipulated, when the contextual state of the phenomenon study is relevant, and also when the boundaries are not clear between this phenomenon and the context. These features are very linked to my research as I try to find out how China's help has affected the economic development of Comoros, European Journal of Business and Management www.iiste.org ISSN 2222-1905 (Paper) ISSN 2222-2839 (Online) Vol.12, No.27, 2020 145 which is a very relevant phenomenon to study.
With this in mind, I will employ a single research draft of the Case Study where the identifiable single case will be China, while the embedded entity in my research will be the Comoros. My choice of the only case lies in China's main role in providing foreign aid to African countries. After comparing the modern foreign aid index of the world's leading aid actors, I got to realize that China was one of the most plausible actors to investigate. With some prior research, I further realize that what can best serve my study and purpose is to find a particular country (embedded entity) that has received this foreign aid from China. When I started brainstorming about which country to choose from the many countries that receive foreign aid from China, it became almost impossible to choose a single country.
With the above criteria in mind, Comoros have made the above criteria very tailor-made. Moreover, the country's exogenous development strategy has contributed to its frenzied diplomatic relations with China, which has been limited by few official visits attended by top officials of both countries.
Required Data Input
With this in mind, my potential data sources, which led to my research on the above indicators, are both secondary and primary sources. Given that there are no secondary data that focus on China's capacity-building assistance, my main sources of secondary education are afforded official document analyses and mapping Chinese development aid in Africa and the European Development Research and Education Association (EADI) report on China's new presence in Africa. The accuracy and reliability of data/information from these sources is not managed by their source, as there is no possibility of this. But since they usually use formal and scientific materials in an academic environment, I trust them.
As for the primary sources that are the main data source for this study, I interviewed nine different outstanding players from different sectors of the Comoros by e-mail. Although the Comoros are a culturally diverse country based on my personal experience and personal observation, these nine interviewees largely represent the greater perception of most Comoros. This is because the interviewees were selected from key sectors, including the Comoros government official, opposition party, business, consultants, banking and financial institutions, national regulatory authorities and civil society.
But with such a drawback, I decided to conduct an interview with an email half-structured open questions that gives the interviewee space to fully formulate their answers. Given that there is a lot of criticism of the data obtained from interviews via email, such as the interviewer's not being able to read facial expressions and body language, make eye contact or hear the voice of the tone of the interviewee (visual and non-verbal signals), Denscombe (2003, p51) recognizes that the quality of response achieved in online studies is the same as traditional methods. He said the same results were achieved in several studies that compared or conducted both email and face-to-face interviews. He also claimed that the interviewees in an email focused more on interview questions and offered more reflective close answers to their colleagues face-to-face. But that doesn't mean that the characteristics of face-to-face interviews are lower, but highlight the benefits of email interviews that give the interviewee time to be more thoughtful and careful about their responses compared to going to an interview with face-face. The analysis of evidence remains one of the least developed and difficult aspects of case studies (Yin, 2003). Marshal and Rossman (2011) assume that bringing order, structure and interpretation to the mass of collected data is messy, ambiguous, time consuming, creative, exciting and does not work linearly. However, I am primarily European Journal of Business and Management www.iiste.org ISSN 2222-1905(Paper) ISSN 2222-2839(Online) Vol.12, No.27, 2020 dependent on categorical aggregation techniques that offer linear and organized analyses of the collected data. It was the first postulate by Robert Stake (1995) as a technique used in case analysis, in which a scientist tries to collect case data in hopes of interpreting the relevant meaning of these cases. In addition, Creswell (2007) describes it as an informative method of drawing meanings across multiple instances of data.
To continue the data analysis, I design 4 categories representing the 4 main themes through which the data from the conversation is encoded by the indicators. These 4 categories are based on an understanding of the likely outcome of China's capacity building through the prism of political realism and interdependence. Forged by the notion that the politically realistic consideration of China's assistance to Africa is oriented toward Chinese interests, as explained above, the likely outcome of China's support for increasing job creation and technology transfer in the Comoros is irrelevant. While, on the other hand, forged the notion that interdependence theorists considered China's assistance to Africa for the mutual benefit of both sides, the likely outcome of China's assistance to increase job creation and technology transfer is significant.
However, it is important to pay attention to important and important averages, such as those used above. It is clear that Comoros are committed to employment opportunities and technology transfers because of the impact of these magnificent support projects in the Comoros. But the biggest question is how big or important these changes are worth paying attention to, which affects economic development. In this context, as is essentially used in the possible results of interdependent theorists, the changes within these two indicators show that they are worth paying sufficient attention to. The pointlessness used to describe the possible consequences of political realist theorists represents a change within these two indicators that is not worth considering. This perception as defined above derives from politically realistic ideology. Unlike Topic, the ideology of interdependence, tries to find evidence within the data that depict china's help that it has enough jobs for the Comoros, which is worth considering. The third theme, formulated from a politically realistic ideology, looks to find evidence within the data that depict China's assistance as not having led to enough ICT transfer, which is worth considering. While other ideology of theoretical dependency, trying to find evidence within the data depicting China's assistance as it leads to enough technology transfer, this is worth considering.
As outlined above, the aim is to discover the kind of impact china has had in the Comoros (create jobs and transfer technology). I will try to draw cases from the data that I can evaluate in the context of analytically oriented indicators. It will greatly help to develop a converging line of evidence that will make my conclusions as strong as possible.
Creation of Jobs: Findings and discussion
Persistent poverty, increase income inequality and slow job growth further exacerbated by the ongoing financial and economic crisis faced by the traditional aid donors, are critical constraints on the socio-economic development of most African countries. Promoting job growth has and continuous to be a central challenge for most of the African countries. With the traditional donors faced with financial and economic crises, China has forged its way into the donor community by providing huge Project aid to the bulk of the African countries. However, with such grandiose projects, one will overwhelmingly expect a job booster within the recipient African countries. Though, in some cases such as the case of Sudan, these grandiose projects have booster employment creation. it was discovered that local residents only provided labour on a marginal scale while a substantial component of the 2222-1905(Paper) ISSN 2222-2839(Online) Vol.12, No.27, 2020 workforce are Chinese. What thus the case of Comoros presents?
ICT Transfer: Findings and Discussion
The importance of technology transfer for economic development cannot be exaggerated. Acquiring technological skills and spreading them in the recipient countries, stimulating economic growth. This would be reflected in the case of Sudan, Uganda and Angola, which are examples of Africans who are best able to describe and assess their support relationship with China, have been largely successful in the field of technology transfer. With Chinese project support, the availability of physical capital (such as machinery and equipment, etc.) and raw materials has increased. In addition, a significant impact on the transfer of technology is the increase in Chinese education (internal and external) programs in these countries. Many citizens working on these Chinese aid projects will be trained simultaneously during the process of implementing Chinese aid and development assistance.
Summary
The above findings and analytical discussion in relation to China's aid towards Capacity building in Comoros, seems to be special. While other Africa countries boost of the positive impact of China's aid towards capacity building in their respective countries as anticipated by the Interdependency theorist, same cannot be said of Comoros. The following table provides a summary of findings. It would be agreed from the table above that, according to the data, the impact of Chinese aid on capacity building in the Comoros was insignificant in light of job creation and ICT transfer. This finding strongly confirms the politically realistic ideology of the potential impact of Chinese assistance to the Comoros. As a portrait on the table above, the interviews gave very strong arguments. On the other hand, the arguments of respondents who considered that this assistance led to a significant increase in job creation and technology, as defined in topic 2 and 4, were weak in transmission. These findings nullify both hypotheses that lead to research. Aware that both hypotheses assumed China's assistance to the Comoros would lead to the creation of employment and technology transfer, this was not the case. Moreover, these findings not only nullify our hypothesis, but also provide us with a very clear answer to the research question about how China's support affects the Capacity building of the Comoros in terms of job creation and ICT transfer.
Conclusion
Indeed, it seems justified to say that few relationships have been so controversial, and still so promising and glued with potentials as that which glues China and Africa together. In many ways, the China -African relationship has raced through both high and low points -the continued use of Chinese workers, the unbalanced nature of the burgeoning trade ties continues to loom the psyche of Africans. Yet in many ways China and Africa need each other. For example, Health centers across most of the Comoros consist of sparsely equipped, cement block rooms. China regularly sends teams of doctors to the country, and it's building a $12 million, modern, eight-story hospital. The Chinese are paving roads, building new schools, new mosques, new government buildings, a new airport, a center to facilitate tourism to the country and even new homes --for politicians. Local TV and radio stations were designed and built by China, and four Chinese government-run television channels now pipe programming from Beijing directly into Comorans' homes. Late last year, China announced a $2 million gift to the Comoros to build a new sports stadium. They all amount to very visible reminders of the nation's benefactors about 6,000 miles away in Beijing.
But the brief of this research was not only to examine the impact of Chinese aid FDI on job creation and ICT transfer, but research why the Comoros case was "exceptional" when most other African countries such as Angola, Sudan and Uganda all have experienced some impetus toward capacity building due to their relationship with China. From the interview, it is clear that corruption and the lack of an organization may be an internal reason why Comoros still finds a brief end to the development staff. In all, the study illustrates some limited Chinese exploration -the Chamber partnership in the light of the power building. Limited by the scope and acquisition of information, the investigation has only acquired highly relevant concerns about employment creation and technology transfer. I just request that I enrich the students and those who are interested in exploring some interesting complexities that reduce this partnership. I hope that wet Buddhist hunger, in the end, can recruit more minds dedicated to asking real relevant questions about China -chamber relations, which have been less explored so far | 6,556.2 | 2020-09-01T00:00:00.000 | [
"Economics"
] |
Properties of charmed and bottom hadrons in nuclear matter: A plausible study
Changes in properties of heavy hadrons with a charm or a bottom quark are studied in nuclear matter. Effective masses (scalar potentials) for the hadrons are calculated using quark-meson coupling model. Our results also suggest that the heavy baryons containing a charm or a bottom quark will form charmed or bottom hypernuclei, which was first predicted in mid 70's. In addition a possibility of $B^-$-nuclear bound (atomic) states is briefly discussed.
Extensive studies with hypernuclei have been carried out over the last 20 years [1,2].These involve embedding a Λ-particle (hyperon), with one (or two) strange quark (quarks) combined with u and (or) d quarks (quark), in finite nuclei and then studying the single particle states, spin-orbit interaction and finally the overall binding of the particle in nuclei with different A, number of ordinary baryons, nucleons, n and p.Such studies have been hindered since there has been no high intensity source of kaon beams that interact with nuclei to produce Λ-particles.
Recently theoretical studies have been extended to take account of the quark structure of the baryons [3,4,5].Agreement with sparse experimental data [2] is impressive.Lately there have been attempts to look for a bound state of 6-quarks, the so-called H particle predicted by Jaffe [6], with no success [7].There has been confirmation of a bound state of two Λ-particles to a finite nucleus (double hypernucleus) [8].All these experimental and theoretical studies were directed to learn about the hadrons containing strange quarks in a surroundings of nuclear sea made of mainly valence u and d quarks, although probably there are no quark studies for the double hypernucleus up to now, in spite of its importance and recent experimental achievements.
The approved construction of the Japan Hadron Facility (JHF) will be essentially a kaon factory, thus it is expected to produce large fluxes of hyperons that should allow a detailed study of hypernuclei.However, the facility will be much more than a kaon factory.With a beam energy of 50 GeV, it will produce charmed hadrons profusely and bottom hadrons in lesser numbers but still with an intensity that is comparable to the present hyperon production rates.In mid 70's, a possible formation of the charmed hypernuclei were predicted theoretically [9,10].There was an experimental search of the charmed and bottom hypernuclei at the ARES facility [11], and it was also investigated at the possible cτ -factory [12].It is clear that situation for the experiments to search for such charmed and bottom hypernuclei is now becoming realistic and would be realized at JHF.This brings us to initiate a careful study of nuclei with charm or bottom quarks.The production of charmonium (cc), mesons with charm, and baryons with charm quarks will be sufficiently large to make it possible to study charmed hypernuclei.Study of such nuclei would initially involve single particle energies, spin-orbit interaction and overall binding energies.Studies with a charm quark and a bottom quark in a many-body system would provide the first opportunity to learn about the behavior of hadrons containing heavy quarks in a sea of valence u and d quarks.Eventually a study of the decay of such hadrons will be a valuable lesson in finding the effect of many-body systems on the intrinsic properties of charmed and bottom hyperons.The advantage of using hadrons with heavy quarks is that they can convey an information at short distance, i.e., that of the very central region of the nucleus from charmed and bottom hypernuclei.Meson nuclear atomic bound states provide useful information about the surface of the nucleus.
The present investigation is devoted to a study of baryons (and mesons) which contain a charm or a bottom quark (will be denoted by C) in nuclear matter.Although the baryons with a charm or a bottom quark which we wish to study have a typical mean life of the order 10 −12 seconds (magnitude is shorter than hyperons), we would like to gain an understanding of the movement of such a hadron in its nucleonic environment.This would lead to an effective mass (scalar potential) for the hadron.The light quark in the hadron (and nucleons) would change its property in nuclear medium in a self-consistent manner, and will thus affect the overall interaction with nucleons.With this understanding we will be in a better position to learn about the hadron properties with the presence of heavy quarks, or baryons with heavy quarks in finite nuclei that will be the real ground for these experimental studies.
At JHF, in addition to charmed and bottom hyperons, mesons with open charm (bottom) like D − (cd) (B − (ūb)) will be produced.Such mesons like K − (ūs) can form mesic atoms around finite nuclei.The atomic orbits will be very small and will thus probe the surface of light nuclei and will be within the charge radii for heavier nuclei.Thus at least for light nuclei they will give a precise information about the charge density.
Furthermore, in considering recent experimental situation on high energy heavy ion collisions, to study general properties of heavy hadrons in nuclear medium is useful, because elementary hadronic reactions occur in high nuclear density zone of the collisions, and many hadrons produced there are under effects of a surrounding nuclear medium.Thus, we need to understand the properties of heavy hadrons in nuclear medium.Some such applications were also made for J/Ψ dissociation in nuclear matter, and D and D productions in antiprotonnucleus collisions [13].
At present we need to resort to a model which can describe the properties of finite nuclei as well as hadron properties in nuclear medium based on the quark degrees of freedom.Although some studies for heavy mesons with charm in nuclear matter were made by QCD sum rule for J/Ψ [14,15] and D(D) [16] there seems to exist no studies for heavy baryons with a charm or a bottom quark.With its simplicity and applicability, we use quark-meson coupling (QMC) model [17], which has been extended and successfully applied to many problems in nuclear physics [18,19,20,21,22,23,24] including a detailed study of the properties of hypernuclei [3], and harmonic properties in nuclear medium [13,25,26,27].In particular, recent measurements of polarization transfer performed at MAMI and Jlab [28] support the medium modification of the proton electromagnetic form factors calculated by the QMC model.The final analysis [29] seems to become more in favor of QMC, although still error bars may be large to draw a definite conclusion.This gives us confidence that such a quark-meson coupling model will provide us with valuable glimpse into the properties of charmed-and bottom-hypernuclei.
We start to consider static, (approximately) spherically symmetric charmed and bottom hypernuclei (closed shell plus one heavy baryon configuration) ignoring small nonspherical effects due to the embedded heavy baryon.We adopt Hartree, mean-field, approximation.In this approximation, ρNN tensor coupling gives a spin-orbit force for a nucleon bound in a static spherical nucleus, although in Hartree-Fock it can give a central force which contributes to the bulk symmetry energy [18,19].Furthermore, it gives no contribution for nuclear matter since the meson fields are independent of position and time.Thus, we ignore the ρNN tensor coupling as usually adopted in the Hartree treatment of quantum hadrodynamics (QHD) [30,31].
Using the Born-Oppenheimer approximation, mean-field equations of motion are derived for a charmed (bottom) hypernucleus in which the quasi-particles moving in single-particle orbits are three-quark clusters with the quantum numbers of a charmed (bottom) baryon or a nucleon.Then a relativistic Lagrangian density at the hadronic level [18,19] can be constructed, similar to that obtained in QHD [30,31], which produces the same equations of motion when expanded to the same order in velocity: where ψ N ( r) (ψ C ( r)) and b( r) are respectively the nucleon (charmed and bottom baryon) and the ρ meson (the time component in the third direction of isospin) fields, while m σ , m ω and m ρ are the masses of the σ, ω and ρ meson fields.g ω and g ρ are the ω-N and ρ-N coupling constants which are related to the corresponding (u,d)-quark-ω, g q ω , and (u,d)-quark-ρ, g q ρ , coupling constants as g ω = 3g q ω and g ρ = g q ρ [18,19].(See also Eqs.( 4) and ( 5).)Note that in usual QMC (QMC-I) the meson fields appearing in Eq. ( 1) represent the quantum numbers and Lorentz structure as those used in QHD [31], corresponding, σ ↔ φ 0 , ω ↔ V 0 and b ↔ b 0 , and they are not directly connected with the physical particles, nor quark model states.Their masses in nuclear medium do not vary in the present treatment.For the other version of QMC (QMC-II), where masses of the meson fields are also subject to the medium modification in a self-consistent manner, see Ref. [20].However, for a proper parameter set (set B) the typical results obtained in QMC-II are very similar to those of QMC-I.The difference is ∼ 16 % for the largest case, but typically ∼ 10 % or less.(For the effective masses of the hyperons, it is less than ∼ 8 %.) In an approximation where the σ, ω and ρ fields couple only to the u and d quarks, the coupling constants in the charmed (bottom) baryon are obtained as g C ω = (n q /3)g ω , and g C ρ = g ρ = g q ρ , with n q being the total number of valence u and d (light) quarks in the baryon C. I C 3 and Q C are the third component of the baryon isospin operator and its electric charge in units of the proton charge, e, respectively.The field dependent σ-N and σ-C coupling strengths predicted by the QMC model, g σ (σ) and g C σ (σ), related to the Lagrangian density, Eq. ( 1), at the hadronic level are defined by: where M N (M C ) is the free nucleon (charmed and bottom baryon) mass (masses).Note that the dependence of these coupling strengths on the applied scalar field must be calculated selfconsistently within the quark model [3,18,19].Hence, unlike QHD [30,31], even though g C σ (σ)/g σ (σ) may be 2/3 or 1/3 depending on the number of light quarks in the baryon in free space (σ = 0)1 , this will not necessarily be the case in nuclear matter.
In the following, we consider the system in the limit of infinitely large, uniform (symmetric) nuclear matter, where all scalar and vector fields become constants.Furthermore, under this limit, we may also treat a hadron h embedded in the nuclear matter system, in the same way as that for the charmed (bottom) baryon.(A Lagrangian density for a meson-nuclear sys-tem can be also written in a similar way to that of the charmed (bottom) hypernuclei system, if L C QM C is replaced by the corresponding meson Lagrangian density in Eq.( 1).)Then, the Dirac equations for the quarks and antiquarks in nuclear matter, in bags of hadrons, h, (q = u or d, and Q = s, c or b, hereafter) neglecting the Coulomb force in nuclear matter, are given by (|x| ≤ bag radius) [25,26,27]: The (constant) mean-field potentials for a bag in nuclear matter are defined by V q σ ≡ g q σ σ, V q ω ≡ g q ω ω and V q ρ ≡ g q ρ b, with g q σ , g q ω and g q ρ the corresponding quark-meson coupling constants.The normalized, static solution for the ground state quarks or antiquarks with flavor f in the hadron, h, may be written, , where N f and ψ f (x) are the normalization factor and corresponding spin and spatial part of the wave function.The bag radius in medium for a hadron h, R * h , will be determined through the stability condition for the mass of the hadron against the variation of the bag radius [17,18,19] (see Eq. ( 8)).
The eigenenergies in units of 1/R * h are given by, The hadron masses in a nuclear medium m * h (free mass will be denoted by m h ), are calculated by where ), and x q,Q being the bag eigenfrequencies.B is the bag constant, n q (n q) and n Q (n Q ) are the lowest mode quark (antiquark) numbers for the quark flavors q and Q in the hadron h, respectively, and the z h parametrize the sum of the center-of-mass and gluon fluctuation effects and are assumed to be independent of density.Concerning the sign of m * q in nuclear medium, it reflects nothing but the strength of the attractive scalar potential as in Eqs. ( 4) and ( 5), and thus naive interpretation of the mass for a (physical) particle, which is positive, should not be applied.The parameters are determined to reproduce the corresponding masses in free space.We chose the values, (m q , m s , m c , m b ) = (5,250,1300,4200) MeV for the current quark masses, and R N = 0.8 fm for the bag radius of the nucleon in free space.The quark-meson coupling constants, g q σ , g q ω and g q ρ , are adjusted to fit the nuclear saturation energy and density of symmetric nuclear matter, and the bulk symmetry energy [17,18,19].Exactly the same coupling constants, g q σ , g q ω and g q ρ , are used for the light quarks in the mesons and baryons as in the nucleon.
However, in studies of the kaon system, we found that it was phenomenologically necessary to increase the strength of the vector coupling to the non-strange quarks in the K + (by a factor of 1.4 2 , i.e., g q Kω ≡ 1.4 2 g q ω ) in order to reproduce the empirically extracted K + -nucleus interaction [25].This may be related to the fact that kaon is a pseudo-Goldstone boson, where treatment of the Goldstone bosons in a naive quark model is usually unsatisfactory.We assume this, g q ω → 1.4 2 g q ω , also for the D, D [27], B and B mesons to allow an upper limit situation.The scalar (V h s ) and vector (V h v ) potentials felt by the hadrons h, in nuclear matter are given by, where I h 3 is the third component of isospin projection of the hadron h.Thus, the vector potential felt by a heavy baryon with a charm or bottom quark, is equal to that of the hyperon with the same light quark configuration in QMC.
In Figs. 1 and 2 we show ratios of effective masses (free masses + scalar potentials) versus those of the free particles, for mesons and baryons, respectively.With increasing density the ratios decrease as usually expected, but decrease in magnitude is from larger to smaller: hadrons with only light quarks, with one strange quark, with one charm quark, and with one bottom quark.This is because their masses in free space are in the order from light to heavy.Thus, the net ratios for the decrease in masses (developing of scalar masses) compared to that of the free masses becomes smaller.This may be regarded as a measure of the role of light quarks in each hadron system in nuclear matter, in a sense that by how much ratio do they lead to a partial restoration of the chiral symmetry in the hadron.In Fig. 1, one can notice somewhat anomalous behavior of the ratio for the kaon (K).This is related to what we meant by the pseudo-Goldstone boson nature, i.e., its mass in free space is relatively light, m K ≃ 495 MeV, and the ratio for the reduction in mass in nuclear medium is large.
Perhaps it is much more quantitative and direct to compare scalar potentials of each hadron in the nuclear matter.Calculated results are shown in Fig. 3. From the results it is confirmed that the scalar potential felt by the hadron h, V h s , follows a simple light quark number scaling rule: where n q (n q) is the number of light quarks (antiquarks) in the hadron h, and V N s is the scalar potential felt by the nucleon.(See Eq.( 9).)It is interesting to notice that, the baryons with a charm and a bottom quark (Ξ c is a quark configuration, qsc), shows very similar features to those of hyperons with one or two strange quarks.Then, we can expect that these heavy baryons with a charm or a bottom quark, will also form charmed (bottom) hypernuclei, as the hyperons with strangeness do.(Recall that the repulsive, vector potentials are the same for the corresponding hyperons with the same light quark configurations.)Thus, an experimental investigation of such hypernuclei would be a fruitful venture at JHF.
In addition, B − meson will also certainly form meson-nuclear bound states, because B − meson is ūb and feels a strong attractive vector potential in addition to the attractive Coulomb force.This makes it much easier to be bound in a nucleus compared to the D 0 [27], which is cū and is blind to the Coulomb force.This reminds us of a situation of the kaonic (K − (ūs)) atom [32,33].A study of B − (ūb) atoms would be a fruitful experimental program.Such atoms will have the meson much closer to the nucleus and will thus probe even smaller changes in the nuclear density.This will be a complementary information to the D − (cd)-nuclear bound states, which would provide us an information on the vector potential in a nucleus [27].
To summarize, we have studied for the first time the properties of heavy baryons (hadrons) which contain a charm or a bottom quark in nuclear matter.Our results suggest that those heavy baryons will form charmed or bottom hypernuclei as was predicted in mid 70's.We plan to report results for the charmed and bottom hypernuclei studied quantitatively, by solving a system equations for finite nuclei embedding a baryon with a charm or a bottom quark [34].In addition we can expect also B − -nuclear bound (atomic) states based on the existing studies for the D 0 and kaonic atom.Furthermore, formation of B − -atoms would provide precise information on the nuclear density, which would be a complementary to that of the D − -nuclear bound states.
Figure 1 :Figure 2 :Figure 3 :
Figure 1: Effective mass ratios for mesons in nuclear matter, where, ρ 0 = 0.15 fm −3 .ω and ρ stand for physical mesons which are treated in the quark model, and should not be confused with the fields appearing in the QMC model. | 4,335.8 | 2002-07-12T00:00:00.000 | [
"Physics"
] |
Bacterial disease induced changes in fungal communities of olive tree twigs depend on host genotype
In nature, pathogens live and interact with other microorganisms on plant tissues. Yet, the research area exploring interactions between bacteria-fungi and microbiota-plants, within the context of a pathobiome, is still scarce. In this study, the impact of olive knot (OK) disease caused by the bacteria Pseudomonas savastanoi pv. savastanoi (Psv) on the epiphytic and endophytic fungal communities of olive tree twigs from three different cultivars, was investigated in field conditions. The ITS-DNA sequencing of cultivable fungi, showed that OK disease disturbs the resident fungal communities, which may reflect changes in the habitat caused by Psv. In particular, a reduction on epiphyte abundance and diversity, and changes on their composition were observed. Compared to epiphytes, endophytes were less sensitive to OK, but their abundance, in particular of potential pathogens, was increased in plants with OK disease. Host genotype, at cultivar level, contributed to plant fungal assembly particularly upon disease establishment. Therefore, besides fungi - Psv interactions, the combination of cultivar - Psv also appeared to be critical for the composition of fungal communities in olive knots. Specific fungal OTUs were associated to the presence and absence of disease, and their role in the promotion or suppression of OK disease should be studied in the future.
www.nature.com/scientificreports www.nature.com/scientificreports/ for studying bacterial-fungal interactions because they have already been shown to provide a special niche for studying microbial multispecies interactions 14 . Indeed, recent studies indicated that some non-pathogenic bacteria, namely Erwinia toletana, Pantoea agglomerans and Erwinia oleae, are frequently associated with Psv in olive knots and effectively cooperate with the pathogen for increasing disease severity 15,16 . Compared with bacterial communities, the fungal community composition in olive knots remains unknown, as well as the way the bacterial pathogen interacts and impacts this fungal community. Using this model system, the simultaneous study of interactions occurring within members of epiphytic or endophytic microbial communities is possible, due to the recognized ability of Psv to live as an epiphyte or endophyte in the olive phyllosphere 12 . Interactions occurring within epiphytic members are of particular interest since the infection of olive tree is believed to be cause by the epiphytic Psv 17 . The availability of olive cultivars with different susceptibility levels to olive knot 12 , which could simultaneously present asymptomatic twigs and knots in the same olive tree, also makes this pathosystem a good model for studying microbial interactions.
With this work, we specifically want to answer the following questions: i) What is the effect of olive knot disease, tree genotype (at cultivar level), and their interaction on fungal communities of twigs? ii) Are these effects identical on epi-and endophytic fungal communities? iii) Is there any fungal consortia associated with olive knot disease and/or host susceptibility? To accomplish this, the composition of both epiphytic and endophytic fungal communities, associated to symptomless twigs and knots from three distinct olive cultivars, were investigated under field conditions. Fungal communities were assessed through PCR identification of culturable isolates. These isolates will be very useful to study the mechanisms of interactions among the most prominent fungi, the pathogen Psv and host plant, and their implication in the control of OK disease. This work is the first step for ascertaining the role of such fungi on OK disease establishment/development in olive tree.
Results
In this study, 179 fungal OTUs belonging to two phyla, 47 families, and 89 genera were identified as inhabitants of olive tree twigs. Ascomycota was the most abundant phylum, accounting for 97.2% of the isolates ( Supplementary Fig. S1). The remaining isolates belonged to Basidiomycota. Fusarium (Nectriaceae), Alternaria (Pleosporaceae), and Cladosporium (Cladosporiaceae) were the most abundant genera, accounting together with 43% of total isolates (Supplementary Fig. S2; Table S1). Indeed, Alternaria was the most frequently isolated in the endophytic community, whereas Cladosporium was dominant in the epiphytic community (Supplementary Fig. S2; Table S1). Epiphytes were significantly more abundant (P < 0.002) and diverse (P < 0.05) than endophytes.
Olive knot disease affects mostly the epiphytic fungal diversity. The diversity of epiphytic and endophytic fungal communities varied significantly between asymptomatic (stems) and OK-symptomatic (knots) twigs, but differences were greater for epiphytes (Fig. 1a). Epiphytic fungi showed a greater decline in abundance (up to 1.7-fold), richness (up to 1.9-fold) and diversity (up to 1.3-fold) in symptomatic twigs (in relation to asymptomatic twigs) than endophytic fungi. In particular, a reduction in abundance of epiphytes belonging to Pleosporaceae, Chaetomiaceae, and Aspergillaceae was evident in symptomatic twigs, and in a lesser extent to Pyronemataceae, Phaeomoniellaceae, Hypocreaceae and Valsariaceae families ( Supplementary Fig. S3). In addition, thirteen epiphytic families disappeared in symptomatic twigs. There was a greater abundance (up to 1.2-fold) of fungal endophytes in OK-symptomatic versus asymptomatic twigs, but a smaller reduction in richness and diversity (up to 1.1-fold) (Fig. 1a). The increase in abundance was mainly due to the Nectriaceae family ( Supplementary Fig. S3), particularly the Fusarium genus (Supplementary Fig. S2; Table S1). Additionally, seven endophytic families, found in asymptomatic twigs, disappeared in symptomatic twigs ( Supplementary Fig. S3).
When we examined fungal guilds, we found the abundance of most epiphytic trophic groups decreased significantly on symptomatic twigs, while the abundance of most endophytic fungi increased (Fig. 1b). Changes in the richness of fungal trophic groups promoted by OK disease were greater for the epiphytic community, where a great decrease in the pathogenic group was observed. Within endophytes, a significant decrease in the richness of beneficial fungi was observed in symptomatic twigs.
Olive knot disease affects mostly the fungal diversity of the most OK-tolerant cultivar. Fungal communities inhabiting asymptomatic twigs varied by cultivar and disease. We found the highest abundance and fungal diversity (P < 0.05) in the most OK tolerant cultivar (cv. Cobrançosa) followed by cvs. Madural and Verdeal Transmontana (Fig. 2a). Fungal communities of each cultivar were differentially affected by OK disease. Knots of cv. Cobrançosa lead to a significantly greater loss of both fungal abundance and richness (up to 1.6-fold, in relation to asymptomatic twigs) than knots in cvs. Madural (up to 1.4-fold) or Verdeal Transmontana (up to 1.1-fold). Most of the lost isolates in knots belonged to the Pezizaceae (for cv. Cobrançosa), Chaetomiaceae (for cv. Madural) and Gnomoniaceae (for cv. Verdeal Transmontana) ( Supplementary Fig. S3). We observed an increase in abundance of Nectriaceae and Pestalotiopsidaceae families in the symptomatic twigs across all cultivars, and an increase in the Mycosphaerellaceae for the most OK-susceptible cultivar (cv. Verdeal Transmontana).
When fungal OTUs were divided into functional categories, host genotype differences were also detected ( Fig. 2b). For all cultivars, a significant decline in abundance and richness of all functional groups was observed in symptomatic twigs in relation to asymptomatic twigs, with the exception of commensals. Abundance of beneficial fungi and richness of pathogens underwent greater declines on cv. Cobrançosa, while pathogenic and commensal fungal abundance, as well as richness of beneficial fungi, were greatly reduced in cv. Madural.
Disease was the major driver of the fungal community composition, whereas host cultivar shaped these communities in symptomatic twigs. As revealed by the fungal clustering in the CCA analysis, the fungal composition in twigs was mainly driven by OK disease rather than by the host genotype (Fig. 3). Results from ANOVA (Table S2) and ANOSIM (Supplementary Table S3) confirm that the presence/absence www.nature.com/scientificreports www.nature.com/scientificreports/ of OK symptoms had the greatest influence on fungal species composition (P = 0.005). Disease explained 3.9%, 5.6% and 5.9% of total, endophytic and epiphytic fungal species variation, respectively; whereas host genotype only explained 0.5%, 1.1% and 1.3% of total, endophytic and epiphytic fungal species variation, respectively (Supplementary Table S2). The effect of host genotype on species composition was greater in symptomatic (P = 0.005) than in asymptomatic (P = 0.049) twigs, explaining 2.8% and 1.5% of species composition variance, respectively (Supplementary Table S2). Differences on fungal species composition between asymptomatic and symptomatic twigs (Supplementary Table S3) were especially noticed in fungal communities of cv. Cobrançosa, in particular for epiphytic community (R = 0.498, P = 0.001). The interaction between OK disease and host genotype was also shown to influence significantly the overall composition of the fungal community (P = 0.010) and the composition of endophytic fungal community (P = 0.005). In contrast, epiphyte composition was not impacted by the interaction between OK disease and host genotype (data not shown).
Specific fungal signatures were detected for the asymptomatic vs. symptomatic twigs. Our previous analyses indicate that OK disease plays an important role on fungal community assemblages, suggesting the existence of a fungal consortium associated to either asymptomatic or OK-symptomatic twigs. To test this hypothesis, a random forest analysis was performed to provide a ranking of the relative importance of each endophyte and epiphyte OTU for distinguishing asymptomatic from OK-symptomatic twigs ( Supplementary Fig. S4). The most distinguishing fungal OTUs were selected and then used to perform a PCA, in order to identify which could be potentially related to OK disease and cultivar (Fig. 4). Among the selected OTUs, some were specifically Figure 1. Comparison of fungal diversity between asymptomatic and OK-symptomatic twigs, either within endophytic or epiphytic communities. (a) Diversity at community level evaluated by determining abundance, richness and by using Shannon-Wiener index. Box plots depict medians (central horizontal lines), the interquartile ranges (boxes), 95% confidence intervals (whiskers), and outliers (black dots). Significant differences between pairs of values are represented over horizontal lines. (b) Changes (%) on fungal abundance and richness for each functional group, occurring on OK-symptomatic twigs in relation to asymptomatic twigs. Asterisks indicate significant differences between these two samples (*P < 0.05, **P < 0.01, ***P < 0.001). (2019) 9:5882 | https://doi.org/10.1038/s41598-019-42391-8 www.nature.com/scientificreports www.nature.com/scientificreports/ associated with a particular set of twigs, either asymptomatic or symptomatic. Heydenia sp., Phaeomoniella sp., Plectania rhytidia, Pyronema domesticum, and Chromelosporium carneum, in the endophytic community, as well as Alternaria alternata, Pyronema domesticum, Hyalodendriella betulae, Epicoccum nigrum, Cladosporium cladosporioides, Coprinellus radians, Neofabraea alba, Trichoderma sp. in the epiphytic community, often occur together and are highly associated with asymptomatic twigs. On the other hand, Fusarium sp., Fusarium lateritium, Biscogniauxia mediterranea, Fusarium oxysporum, Neofabraea alba, Alternaria sp., and Alternaria tenuissima, in the endophytic community, as well as Penicillium spinolosum, Penicillium canescens and Comospora sp., in the epiphytic community, were simultaneously found on symptomatic twigs (knots). However, a clear association of these fungal OTUs and olive cultivars was not found. In order to discover significant associations between fungal OTUs of asymptomatic/symptomatic twigs and olive cultivars, a species indicator analysis was carried out using preselected OTUs by the random forest analysis. The best indicator OTUs (IndVal > 0.70) of asymptomatic and symptomatic twigs were found in cv. Cobrançosa (Table 1). Curiously, H. betulae was identified as an epiphyte indicator of asymptomatic twigs for all three cultivars, while the endophyte Fusarium was found to be a good indicator of symptomatic twigs for all three cultivars.
Figure 2.
Comparison of fungal diversity between asymptomatic and OK-symptomatic twigs within each olive tree cultivar (Cobrançosa, Madural and Verdeal Transmontana). (a) Diversity at community level by determining abundance, richness and by using Shannon-Wiener index. Box plots depict medians (central horizontal lines), the inter-quartile ranges (boxes), 95% confidence intervals (whiskers), and outliers (black dots). Significant differences between pairs of values are represented over horizontal lines. (b) Changes (%) on fungal abundance and richness for each functional group, occurring on OK-symptomatic twigs in relation to asymptomatic twigs. Asterisks indicate statistically significant differences between these two samples (*P < 0.05, **P < 0.01, ***P < 0.001).
Discussion
A great diversity of fungal epiphytes and endophytes (in total 179 OTUs) was found in twigs/knots of olive tree. This fungal community was studied based on cultivation-dependent method and therefore it is expected that twigs/knots harbor a much larger diversity. Indeed, such approach has limitations due to some non-sporulating and non-culturable fungi 18 . Despite the limited coverage of culture-based methods, previous studies have indicated that, the dominant fungal taxa identified by cultivation-independent methods were also detected by culture-dependent methods 19 . In this work, we had used culture-based method because it offers the possibility to isolate and study the fungal strains for their biological features in terms of biological control of olive knot disease and mechanisms involved.
Our comparative community analysis of fungi between asymptomatic and OK-symptomatic twigs showed that OK disease primarily affects the epiphytic fungal community, by decreasing abundance, richness, and diversity, suggesting that Psv may prevent fungal colonization and proliferation on the knot surface. Antifungal activity displayed by other Pseudomonas species has been already reported, and applied in the biocontrol of a wide range of fungal phytopathogens 17 . Pseudomonas have been shown to produce antibiotics and fungal cell wall www.nature.com/scientificreports www.nature.com/scientificreports/ degrading enzymes, as well as a competitor for space and nutrient sources 20 . However, very few studies have explicitly examined the antifungal activity promoted by phytopathogenic Pseudomonas 9 . There is evidence suggesting phytopathogenic Pseudomonas may inhibit the growth and extract nutrients from filamentous fungi in www.nature.com/scientificreports www.nature.com/scientificreports/ the phyllosphere 21 . Thus, Psv may also suppress fungal colonization/proliferation on olive knot surface in a similar manner. Compared to epiphytes, the endophytic fungal diversity and richness in knots was less impacted, suggesting endophytes are less sensitivity to OK disease. But the fungal endophyte abundance increased in olive knots, suggesting an unexpected stimulatory effect of Psv on endophytic fungi, particularly Pseudocercospora spp. (Mycosphaerellaceae family), which are well-recognized plant pathogens on a wide range of plant hosts, including olive trees 22 . This increasing abundance in the presence of Psv could result from the production of specific compounds by the pathogenic bacteria that could benefit endophytic fungi 9 . Effects of Pseudomonas on fungal growth and density stimulation have already been demonstrated in a number of settings, including mushroom formation 23,24 and human infections 25 .
The effect of OK disease on fungal richness and abundance was greater in the OK-tolerant cv. Cobrançosa, compared to both OK-susceptible cultivars. Probably there is a cultivar interaction that promoted the Psv antifungal activity involving cv. Cobrançosa, either due to the cultivar chemical composition or because this cultivar had a different initial fungal community. This effect has been previously observed in medicine 26 . Clinical research on the role of Candida-bacteria interactions in disease indicated that host's microbial community might play an important role in the preparation of the fungus for its role in the infection 26 . In our study, a strong decrease in Chromelosporium carneum (Pezizaceae) abundance was observed, suggesting that this species is highly sensitive to OK disease aspects in the plant. Future work is needed to validate C. carneum inhibition by Psv and to elucidate the importance of C. carneum abundance in the knot habitat.
The decline in both fungal richness and abundance between asymptomatic and symptomatic twigs was primarily driven by less studied fungal OTUs, including the genera Heydenia, Hyalodendriella, Masonia, Ochrocladosporium or Prosthemium (data not shown). These results open the field for exploring an untapped diversity with potential to benefit olive tree health.
Beneficial and pathogenic fungi were affected by OK disease, but in different ways depending on the type of fungal community (endophytic or epiphytic). On the knot surface, pathogenic fungi decreased in abundance in almost the same proportion as beneficial fungi, indicating that OK disease restricted the growth of both functional groups to the same degree. In contrast, in the knot interior, an increase in pathogenic fungal abundance was accompanied by a decrease in beneficial fungal richness. Some of the lost beneficial fungi were members of Epicoccum, Cladosporium, and Penicillium genera that included antagonists and disease-protective fungi [27][28][29] . The decline of these fungal genera in olive knots suggests they can potentially limit or prevent OK disease. Pathogens with increased abundance included mostly non-pathogens of olive tree. The potentially role of these beneficial and pathogenic groups in the development of OK disease must be studied in future works.
The composition of the fungal community of olive twigs was primarily impacted by OK disease. This may be due to the Psv ability to shape and change the shared environment either to the benefit or to detriment of certain fungal species 30 . Indeed, Psv has the ability to form biofilm 31 , which represent a hotspot for microbial interactions that locally shape microbial assemblages 2,9 . Bacterial-Psv interactions have already been reported to occur in olive knots, where the presence of Psv is suggested to be essential for the creation and maintenance of a core group of bacterial genera observed in olive knots 16 . These multispecies interactions observed among bacteria may also occur between Psv bacteria and fungi. In addition, pathogen is likely to trigger host susceptibility, which could have an impact on fungal community composition of olive twigs. Indeed, there are some reports indicating that some pathogenic infections can be detrimental to the defense systems predisposing the plant to subsequent secondary infections 32 . For instances, infection by Pseudomonas syringae was showed to render Arabidopsis plants more vulnerable to invasion by the necrotrophic Alternaria brassicicola 33 .
The cultivar affects mostly the fungal community composition of symptomatic twigs (i.e. knots) compared to the ones of asymptomatic twigs. Thus, the interaction of cultivar and Psv seems to influence the establishment of fungal communities in olive knots, likely because host plants play a role in recruiting fungi upon OK disease establishment, as previously showed in the Arabidopsis thaliana rhizosphere upon foliar pathogen attack 34 . The composition of endophytic fungal community in olive twigs was affected by the interaction between OK disease and cultivar, thus reinforcing the hypothesis of recruitment of fungi by the plant upon Psv attack.
Our data revealed that a consortium of fungal OTUs is associated with asymptomatic or symptomatic twigs of each cultivar. The most parsimonious assumption is that these consortia probably have relevance to olive tree health. The best indicator OTUs of asymptomatic twigs are antagonists toward plant pathogens (C. cladosporioides in cv. Cobrançosa 35 ), and fungi with unknown biological function (e.g. P. domesticum in cv. Madural, C. carneum in cv. Verdeal Transmontana, and H. betulae in the three cultivars). Alternaria sp. was also found to be indicative of asymptomatic twigs in cv. Cobrançosa. The genus Alternaria includes both plant-pathogenic and saprophytic species, and is one of the most well-known fungal genera producers of diverse secondary metabolites, including toxins 36 and antimicrobial compounds 37 .
The best indicator OTUs for olive knots comprise members of Fusarium, a genus including common plant pathogens 38,39 . This taxonomic group has been described as "pathogen facilitators" that may aid pathogen infection of plant hosts or increase disease severity 40 . A strong association was also detected between A. alternata and knots of cv. Madural. This is a generalist saprobe fungal species, but also contains several variants, which cause necrotic diseases on different plants 41 . Fusarium and Alternaria genera include fungal species that have been recently reported to cause olive tree diseases with minor importance 42,43 . The role of these fungal taxa associated to either asymptomatic or symptomatic twigs, on olive tree's defense against OK disease remains a topic for further study.
In summary, our study indicates that OK disease caused by the bacterium Psv alters the resident fungal community of olive twigs in terms of species composition, abundance and richness. This effect was strongest for epiphytes. In the interior of olive knots, Psv seems to shape fungal assemblages possibly by promoting pathogens. Host plant genotype is also likely to structure olive knot-associated fungal communities. Specific fungal signatures were detected for asymptomatic and symptomatic twigs, suggesting an important role of the fungal www.nature.com/scientificreports www.nature.com/scientificreports/ community in OK disease establishment and development. These results represent an important step forward in understanding the complexity of interactions between bacteria-fungi, and host-microbe interactions, which is needed for predicting and suppressing OK disease.
Methods
Plant sampling. Plant collection was performed during spring 2014, in three olive orchards located in Mirandela, Northeast of Portugal, at coordinates N 41° 32.593′; W 07° 07.445′ (orchard 1), N 41° 32.756′; W 07° 07.590′ (orchard 2) and N 41° 29.490′; W 07° 15.413′ (orchard 3). Each orchard comprised olive trees from three cultivars, planted at 7 × 7 m spacing, with different levels of susceptibility to olive knot disease: cv. Cobrançosa is moderately tolerant, cv. Madural is moderately susceptible and cv. Verdeal Transmontana is the most susceptible. Their susceptibility was confirmed by estimating OK disease incidence simultaneously to sample collection. The levels of disease incidence (%), determined by the percentage of infected twigs, were indeed lower in cv. Cobrançosa (7.3 ± 3.2) when compared to cvs. Madural (13.4 ± 6.7) and Verdeal Transmontana (17.2 ± 9.8). Twigs were sampled from randomly selected seven olive trees of each cultivar, from which asymptomatic and OK-symptomatic (with 3-4 knots) twigs were collected from two cardinal orientations (north and south), at 1.5-2 m above ground height.
Fungal isolation. From each tree, stem and knot segments (N = 5 each, with around 1 gram of weight/each) were randomly selected from asymptomatic and symptomatic twigs, respectively. These plant tissues were used for isolation of fungal epiphytes and endophytes. Fungal epiphytes were isolated by the plate dilution method, onto Potato Dextrose Agar (PDA, Difco) and Plate Count Agar (PCA, Himedia) media, supplemented with 0.01% (w/v) chloramphenicol (Oxoid), following the procedure described by Gomes et al. 44 . The number of epiphytes was expressed as log CFU/cm 2 , i.e. the number of individual colonies of fungi adhered to stem/knot surface. To estimate plant tissues surface, a cylinder equation [A = 2πr 2 + h (2πr)] was used, in which A is the area, r and h are the radius and height of the stem/knot segments, respectively. The average stem/knot segments area of cvs. Cobrançosa, Madural and Verdeal Transmontana were 11.2 ± 0.9, 10.8 ± 1.7 and 11.9 ± 1.8 cm 2 , respectively.
Endophytic fungi were isolated from the same stem/knot segments used to isolated epiphytes, following the procedure described by Martins et al. 45 . Briefly, after surface disinfection, each stem/knot was cut in segments (ca. 4-5 mm), which were transferred to the same culture media used for epiphyte isolation. Validation of the surface sterilization procedure was done by imprinting the surface of sterilized twigs onto PDA and PCA media. A total of 7,440 plant tissue segments were used to isolated fungal endophytes. Fungal colonies were subcultured on fresh medium until pure epi/endophytic cultures were obtained.
Fungal identification. We first separated fungal isolates based on their morphological features (both colony shape and microscopic morphology). Then we selected three representative isolates of each morphotype for molecular identification by sequencing the internal transcribed spacer (ITS) region of nuclear ribosomal DNA (rDNA). Total genomic DNA was extracted from harvested mycelium or spores using the REDExtract-N-Amp ™ Plant PCR kit (Sigma, Poole, UK). The ITS region (ITS1, 5.8S, ITS2) was amplified using ITS1/ITS4 or ITS5/ ITS4 primers sets 46 in a PCR protocol previously described by Oliveira et al. 47 . The amplified products (~650 bp) were purified and sequenced using Macrogen Inc. (Seoul, South Korea) services. The obtained DNA sequences were analyzed with DNASTAR v.2.58 software, and fungal identification was performed using the NCBI database (http://www.ncbi.nlm.nih.gov) and BLAST algorithm 44 . The obtained sequences are available at GenBank with the following accession numbers: KU324941-KU325040; KU325041-KU325240; KU325241-KU325457. Each operational taxonomic unit (OTU) was taxonomically classified according to the Index Fungorum Database (www.indexfungorum.org). Pure cultures of each identified isolate were preserved and deposited in the culture collection of the Polytechnic Institute of Bragança (School of Agriculture).
Effect of OLS disease and host genotype on fungal diversity. The effect of OK disease or host genotype in twig fungal diversity was analyzed at both community and functional group levels. We evaluated community level diversity by determining the abundance (number of isolates), richness (number of different OTUs), and computing the Shannon-Wiener (H') index in Species Diversity and Richness v. 4.0 48 . Results are presented as the mean of replicates (N = 21, for each cultivar). We evaluated functional group by placing identified fungi into functional categories (commensal, beneficial and pathogenic) according to the description of fungal endophytes by Hardoim et al. 49 . The commensal group is comprised of fungi that do not have any apparent effect on host plants, whereas beneficial fungi can protect host plants against pathogens and pests and/or promote plant growth. The pathogen group includes latent plant pathogens. Fungal OTUs belonging to Incertae sedis or other functional groups, were categorized as "unknown fungi" or "other", respectively. The several taxa were placed in functional categories using expert knowledge (Table S4). After grouping, the relative abundance and richness of each functional group across asymptomatic and OK-symptomatic twigs was determined. To determine differences in fungal community or functional diversity among twigs, a one-way analysis of variance (ANOVA) with SPSS v.20 was performed, with multiple comparisons according to Tukey (P < 0.05).
Data analysis.
A combination of univariate and multivariate methods were used to identify potential fungal OTUs that differentiate olive tree cultivars (cvs. Cobrançosa, Madural and Verdeal-Transmontana) and twig status (asymptomatic and symptomatic). All statistical analyses were performed in R 50 .
Effect of OK disease and host genotype on fungal community structure. A canonical correlation analysis (CCA) was used to find possible correlations among surveyed cultivars or twig status with the identified fungal communities (endophytes, epiphytes and total). All data were log 2 (x + 1) transformed for standardization. The CCA was performed using the "CCorA" function in the vegan package 51 . One-way analysis of variance (ANOVA) www.nature.com/scientificreports www.nature.com/scientificreports/ was performed with the "anova" function to determine statistical significant differences among cultivars and twig status, on endophytic, epiphytic and total fungal communities. Variation partitioning was used to calculate community dissimilarity (%), according to different explanatory variables (olive cultivar and presence/absence of OK-symptoms). This analysis was also performed with the vegan package using the "varpart" function 52 . A one-way analysis of similarity (ANOSIM) was used to test statistic differences between fungal groups separated by the CCA, using the Bray-Curtis distance matrices. This analysis was performed using the "anosim" function in the vegan package 51 .
Identification of fungal OTUs associated with OK disease and/or host genotype. A random forest analysis using artificial intelligence algorithms was performed to identify the ranking importance of fungal OTUs for distinguishing asymptomatic from OK-symptomatic twigs 53,54 . For each tree grown on a bootstrap sample, the error rate for observations left out of the bootstrap sample was monitored. The mean decrease in Gini coefficient indicates the variable contribution for distinguishing between asymptomatic and OK-symptomatic twigs. The ranking species importance explained 81.4% and 84.2% for endophytes and epiphytes, respectively. We then conducted a principal component analysis (PCA) and indicator fungal species analysis using the species pre-selected by the random forest analysis. Both analyses were used to explore the potential associations between fungal OTUs and cultivar/twig status. The PCA was performed by using the psych package 55 . The indicator fungal species analysis was conducted using the function "multipatt" from indicspecies package 56 . We used the Indicator Value Index (IndVal) (IndVal > 0.5 represents the most constant and specific species) which is defined as the product of two components: "A", the specificity of the species as indicator of the site group and "B"; the sensitivity of the species as indicator of the site group 56 . | 6,361.6 | 2019-04-10T00:00:00.000 | [
"Environmental Science",
"Biology"
] |
ARAP3 Functions in Hematopoietic Stem Cells
ARAP3 is a GTPase-activating protein (GAP) that inactivates Arf6 and RhoA small GTPases. ARAP3 deficiency in mice causes a sprouting angiogenic defect resulting in embryonic lethality by E11. Mice with an ARAP3 R302,303A mutation (Arap3KI/KI) that prevents activation by phosphoinositide-3-kinase (PI3K) have a similar angiogenic phenotype, although some animals survive to adulthood. Here, we report that hematopoietic stem cells (HSCs) from rare adult Arap3KI/KI bone marrow are compromised in their ability to reconstitute recipient mice and to self-renew. To elucidate the potential cell-autonomous and non-cell-autonomous roles of ARAP3 in hematopoiesis, we conditionally deleted Arap3 in hematopoietic cells and in several cell types within the HSC niche. Excision of Arap3 in hematopoietic cells using Vav1-Cre does not alter the ability of ARAP3-deficient progenitor cells to proliferate and differentiate in vitro or ARAP3-deficient HSCs to provide multi-lineage reconstitution and to undergo self-renewal in vivo. Thus, our data suggest that ARAP3 does not play a cell-autonomous role in HSPCs. Deletion of Arap3 in osteoblasts and mesenchymal stromal cells using Prx1-Cre resulted in no discernable phenotypes in hematopoietic development or HSC homeostasis in adult mice. In contrast, deletion of Arap3 using vascular endothelial cadherin (VEC or Cdh5)-driven Cre resulted in embryonic lethality, however HSCs from surviving adult mice were largely normal. Reverse transplantations into VEC-driven Arap3 conditional knockout mice revealed no discernable difference in HSC frequencies or function in comparison to control mice. Taken together, our investigation suggests that despite a critical role for ARAP3 in embryonic vascular development, its loss in endothelial cells minimally impacts HSCs in adult bone marrow.
Introduction
Hematopoietic stem cells (HSCs) are the critical source of all blood cells. Their potential for self-renewal and multi-lineage repopulation sustains the rapid turnover of the blood system throughout life. The first HSC arises from the hemogenic endothelium in the Aorta-Gonad-Mesonephros (AGM) region of the embryo and subsequently colonizes the fetal liver [26]. In the adult mouse, HSCs reside in complex bone marrow (BM) niches that are not mutually exclusive. Extensive research has shown that HSC perivascular and osteoblastic niches are comprised of endothelial cells, mesenchymal stromal cells, osteoblasts, sympathetic nerves and non-myelinating Schwann cells [1][2][3].
ARAP3 is a dual Arf and Rho GTPase-activating protein that was first identified in porcine leukocytes for its ability to bind to phosphatidylinositol (3,4,5)triphosphate (PIP 3 ) [22]. ARAP3 contains two distinct GAP domains that accelerate the rate of GTP hydrolysis to attenuate Arf6 and RhoA signaling [23,24]. Previous in vitro studies found that either exogenous ARAP3 expression in epithelial cells or RNAi-mediated ARAP3 depletion in endothelial cells disrupts F-actin or lamellipodia formation, respectively, resulting in a cell rounding phenotype and failure to spread [25,26]. This implies that ARAP3 controls Arf6 and RhoA in a tightly regulated fashion, and that maintaining precise regulation of ARAP3 activity is crucial to actin organization in the cell. RhoA has been characterized in vivo to regulate migration and chemotaxis of mature hematopoietic cells [27,28], as well as HSPC engraftment, multi-lineage repopulation and cell survival [9,14,15], while the role of Arf6 in hematopoiesis is largely unknown.
In mice, ARAP3 is most highly expressed in the endothelium and bone marrow, and has been found to be critical to vascular development [29,30]. Germline deletion or Tie2-Cre-mediated deletion of Arap3 in mice leads to embryonic lethality by E11 due to defects in sprouting angiogenesis of the endothelium [29]. Since HSCs arise from the hemogenic endothelium during embryonic development around E10.5 [31], and give rise to all subsequent hematopoietic cells in the fetal liver and in the adult BM, this genetic model precludes further studies of ARAP3 function in definitive hematopoiesis and HSC function. Conditional Arap3 deletion in neutrophils has been shown to alter their adhesion-dependent functions [32,33], but the role of ARAP3 in HSPCs has yet to be defined.
ARAP3 is a phosphoinositide 3-OH kinase (PI3K)-and Rap-regulated GAP that is recruited to the plasma membrane in a PIP 3 -dependent fashion. PI3Kdependent activation of ARAP3 involves binding of its two most N-terminal pleckstrin homology (PH) domains to PIP 3, a lipid second messenger generated downstream of PI3K. This drives recruitment of ARAP3 to the plasma membrane to facilitate interaction with its GTPase substrates [34]. PIP 3 binding is ablated when a tandem arginine to alanine mutation is introduced at residues R302,R303 in the first PH domain of ARAP3, preventing ARAP3 activation and recruitment to the plasma membrane [29]. Arap3 R302,303A/R302,303A knock-in mutant mice (here referred to as KI/KI) phenocopy Arap3 null mice, suggesting an essential role for PI3K-dependent activation of ARAP3 [29].
In this study, we first investigate ARAP3 function in adult hematopoiesis using KI/KI 2 mice, since about 2% of KI/KI mice are viable [29,33]. We report that KI/ KI HSCs are compromised in their ability to repopulate and self-renew in serial transplantation assays. To elucidate potential cell-autonomous and non-cellautonomous roles for ARAP3 in HSC function, we selectively delete Arap3 in the hematopoietic compartment, and in endothelial and stromal cells of the HSC niche, respectively, using Cre driven by suitable promoters (Vav1, Prx1, Cdh5/VEcadherin). Using these genetic models, we report that ARAP3 does not play a major role in regulating HSPC functions.
Arap3 R302,303A mutation impairs HSC functions
To study whether ARAP3 function affects adult hematopoiesis and HSC function, we studied KI/KI mice expressing mutant ARAP3 R302,303A. This point mutation interferes with the ability of ARAP3 to bind PIP 3 and its subsequent activation by PI3K [29,33]. Most KI/KI mutant mice die embryonically at E11 [29], but a small subset (,2%) was viable and fertile when the expected birth ratio was 25% (Table 1). By 8-12 weeks of age, these mice were indistinguishable from their littermate controls in gross appearance as well as by phenotypic characterization of their peripheral blood (Fig. 1A). KI/KI BM also showed normal progenitor cell numbers as determined by colony-forming cell (CFC) assays (Fig. 1B), and normal HSC frequencies as determined by flow cytometry using SLAM family surface markers, CD48 2 CD150 + LSK (Lin 2 Sca1 + cKit + ) [35] (Fig. 1C).
To study the function of these mutant HSPCs, purified LSK cells from KI/KI mutant mice or control mice were injected with competitor bone marrow cells into each irradiated recipient mouse. Reconstitution in individual recipient mice was followed every 4 weeks post-transplant. We found that KI/KI LSKs displayed a significantly lower donor chimerism in all lineages of the peripheral blood from recipient mice (Figs. 1D and 1E). Donor-derived cells in the BM, LSK, and SLAM LSK (CD48 2 CD150 + LSK) compartments were significantly lower in mice transplanted with KI/KI cells, in comparison to control cells (Fig. 1F). Four months after the primary transplant, BM cells were harvested and 2610 6 total BM cells were injected into secondary irradiated recipient mice. Tertiary transplants were performed similarly. Peripheral blood and BM HSC reconstitution after each transplant was analyzed by flow cytometry. We found that the defects in reconstitution of KI/KI cells were exacerbated upon serial transplantations, indicating compromised HSC self-renewal (Figs. 1G-1I). Interestingly, a myeloid bias in the multi-lineage reconstitution of serially transplanted recipients arose within KI/KI reconstituted mice (Figs. 1J and 1K), reminiscent of aged HSCs [36][37][38].
ARAP3 is dispensable for steady-state hematopoiesis
Our data from the knock-in mice prompted us to further elucidate the role of ARAP3 in regulating HSC function using ARAP3 conditional knockout (CKO) mice. To study if ARAP3 plays a cell-autonomous role in hematopoietic cells, we crossed Arap3 flox/flox (f/f) mice with mice expressing a Vav1 promoter-driven Cre. Vav1 expression begins around E11.5, and is fully turned on and expressed in greater than 99% of hematopoietic cells by E13.5, thereby excising floxed alleles in most, if not all, fetal liver and adult hematopoietic cells [39].
To measure the deletion efficiency at the DNA level, we cultured 1.5610 4 unfractionated BM cells from either Arap3 flox/flox ;Vav1-Cre tg (f/f;Vav) CKO or f/f control mice in semi-solid methylcellulose cultures. Each progenitor cell gives rise to an individual colony, from which DNA was isolated to measure Arap3 deletion efficiency on a clonal basis. Clonal analysis by polymerase-chain reaction (PCR) genotyping showed near 100% Vav1-Cre-mediated excision of Arap3 (S1A Fig.). We also measured Arap3 deletion efficiency at the transcript level using quantitative real-time PCR (qRT-PCR). f/f;Vav showed a greater than 95% deletion of Arap3 transcripts in the BM when compared to f/f control mice (S1C Fig.). In contrast, the transcript levels of the other ARAP family members, Arap1 and Arap2, remained unchanged in the BM (S1D Fig.). To ensure a more complete deletion of ARAP3 in hematopoietic cells, we generated Arap3 flox/2 ; Vav1-Cre tg (f/2;Vav) mice. These mice showed 100% deletion in all f/2;Vav mice we examined at both DNA and RNA levels ( Figs. 2A and 2B). We found that both
ARAP3 does not play a cell-autonomous role in HSCs
We next examined if ARAP3 affects primitive hematopoietic compartments. We found that f/2;Vav mice had a normal distribution of phenotypic HSPCs within the LSK compartment, as determined by SLAM markers (Fig. 2E). Furthermore, Arap3 deficiency in BM cells did not affect hematopoietic progenitor cell proliferation or differentiation in CFC assays ( Fig. 2F and S2C Fig.). Of note, when purified f/f;Vav LSKs were plated, they exhibited abilities to form various types of colonies with normal frequencies (S2D Fig.) and morphology (not shown). In contrast, Arap3 2/2 neutrophils from the CKO mice exhibited enhanced polyRGD-induced adhesion (S2E Fig.), as previously published [32]. These data together suggest the cell-intrinsic functions of ARAP3 are limited to more differentiated myeloid cells, rather than in the immature HSPC populations. We next assessed whether ARAP3 plays a role in HSC function in vivo using competitive bone marrow transplantation (BMT) assays. LSK cells sorted from f/ f;Vav mice or f/f littermate controls were transplanted into lethally-irradiated recipient mice and peripheral blood reconstitution was evaluated. There were no significant differences in multi-lineage reconstitution (S2H Fig.) or donor chimerism, either in the total blood cell population (S2F Fig.) or in myeloid/Tcell/B-cell lineages after primary BMT (S2G Fig.). Secondary transplants also showed normal multi-lineage reconstitution (S2I-S2K Figs.). Furthermore, serial transplantation of f/2;Vav LSKs showed no discernable differences in donor contribution (Figs. 2G and 2I) or multi-lineage reconstitution (Figs. 2H and 2J) when compared to littermate Arap3 flox/2 (f/2) or f/f control donors. Together, our data firmly established that ARAP3 does not cell-autonomously impact HSC homeostasis or function. We next investigated whether ARAP3 plays a role in the HSC niche, acting noncell-autonomously to regulate HSCs. To study this, we deleted Arap3 using Prx1-Cre, shown to induce excision in nearly all osteoblasts and 95% of perivascular stromal cells but not in endothelial cells [2,40]. Arap3 flox/2 ;Prx1-Cre tg (f/2;Prx1) mice were born alive in expected Mendelian ratios. Hematopoietic development of f/2;Prx1 mice appeared normal as determined by CBC and lineage distribution in hematopoietic tissues, as well as by progenitor numbers and function in CFC assays (data not shown). f/2;Prx1 mice also exhibited normal frequencies of phenotypic HSCs and progenitors as characterized by flow cytometric analysis (Fig. 3A). Using competitive BMT assays, we found that f/2;Prx1 LSKs repopulated comparably to control f/2 LSK cells (Fig. 3B). These data indicate that ARAP3 in Prx1-expressing niche cells is not a significant regulator of steadystate hematopoiesis or HSC homeostasis.
ARAP3 expression in endothelial cells is important for embryonic development but not adult HSC functions
Arap3 deletion in endothelial cells using Tie2-Cre resulted in embryonic lethality due to a cell-autonomous angiogenesis defect, though only with more robust excision on the f/2, but not f/f, background [29]. Since Tie2-Cre also shows some expression in stromal cells [41,42], we utilized the vascular endothelial cadherin (VEC or Cdh5) promoter-driven Cre (VEC-Cre) that is found to be more endothelial cell-specific than Tie2-Cre [43,44]. Arap3 flox/flox ;VEC-Cre tg (f/f;VEC) mice were born at expected Mendelian ratios (Table 2).
To obtain a more complete deletion of Arap3 in endothelial cells, we generated Arap3 flox/2 ;VEC-Cre tg (f/2;VEC) mice. These CKO mice were born at a significantly reduced ratio ( Table 2). The surviving CKO mice appeared grossly normal, and showed an approximate 88% excision efficiency (Fig. 4A), ranging from 80% to 96%, compared to an average 80% excision rate in f/f;VEC mice (S1B Fig.). By qRT-PCR, an approximate 9% Arap3 transcripts remained in f/2;VEC BM ( Fig. 4B), in contrast to 15% in f/f;VEC BM (S1C Fig.). In agreement with previously published data, this indicates that ARAP3 function in endothelial cells is essential for embryonic development.
Both f/f;VEC and surviving f/2;VEC adult mice had similar peripheral blood composition to their control littermates (S3A Fig. and Fig. 4C) and normal lineage distribution in hematopoietic tissues (S3B Fig. and Fig. 4D). These mice also exhibited a normal distribution of immature HSPC populations, with expected numbers of CD48 2 CD150 + LSK cells (Fig. 4E). Hematopoietic progenitor numbers were normal as determined by CFC assays (Fig. 4F and S3C Fig.). We next examined HSC functions in vivo using competitive BMT assays. To assess whether ARAP3 expression in BM endothelial cells is important to support HSC functions, we performed reverse transplantation of wild-type bone marrow cells into f/f and f/2 control (SJL:CTL) or f/f;VEC and f/2;VEC CKO (SJL:VEC-CKO) recipient mice. HSC homeostasis or engraftment was not altered, and multi-lineage reconstitution of the hematopoietic compartment was unaffected by Arap3 deletion in the host endothelial HSC niche ( Fig. 4K and data not shown). Taken together, our investigation suggests that in spite of a critical role for ARAP3 during embryonic vascular development [29,30], loss of ARAP3 in endothelial cells minimally impacts HSCs in the adult BM.
Discussion
In the present study, we utilize KI/KI mice and generate three Arap3 CKO mouse models to study ARAP3 functions in hematopoiesis of the adult mouse. We show that KI/KI mice exhibit defective HSC functions upon transplantation, indicating that PI3K-mediated ARAP3 function is important for HSCs. To elucidate the potential cell-autonomous and non-cell-autonomous roles of ARAP3 in HSCs, we conditionally deleted Arap3 in hematopoietic cells and in several cell types in the HSC niche. Arap3 null HSCs from f/2;Vav mice were functionally competent to repopulate the HSC pool and self-renew upon serial transplantation, revealing ARAP3 is not required in regulating HSC homeostasis or function in a cellautonomous manner. Furthermore, ablation of ARAP3 in perivascular stromal cells, osteoblastic cells, and endothelial cells of the HSC niche did not alter HSC function or maintenance in f/2;Prx1 and f/2;VEC mice. Reverse transplantation experiments strongly suggest that ARAP3 expression in the BM endothelial niche is not required to support HSC functions; however, future investigation is needed to assess the potential role of ARAP3 in BM mesenchymal and osteoblastic niches.
Our study demonstrates that f/2;VEC mice displayed partial embryonic lethality, indicating a critical role for ARAP3 in endothelial cells during embryonic development but not in the adult BM niche for HSCs.
ARAP3 was first purified from porcine leukocytes for its PIP 3 binding ability and is highly expressed in hematopoietic tissues, however our studies show that ARAP3 is not a critical cell-intrinsic regulator of hematopoiesis. ARAP3 does not play a cell-autonomous role in regulating HSC homeostasis or function. However, our observation does not rule out a cell-autonomous role for ARAP3 under stress conditions. This phenomenon has been seen in genetic studies of other proteins, such as SIRT1 [45,46]. It is possible that a physiological challenge to the mice is necessary to elicit a specific function for ARAP3 in hematopoiesis. In our model, Arap3 is excised in hematopoietic cells during early development using Vav1-Cre, and compensatory mechanisms may account for normal hematopoiesis in adult mice. It would be interesting to study whether acute loss of ARAP3 in hematopoietic cells of the adult mouse, such as with an inducible Cre system [47], would reveal a role for ARAP3 in hematopoiesis. Additionally, ARAP3 is part of a dual-GAP family that also includes ARAP1 and ARAP2. Although each family member targets different G-protein substrates and has different active functional domains [48][49][50], these three proteins may have overlapping and redundant roles in hematopoiesis, and may work in conjunction to regulate HSPC function. Thus, genetic studies of mice deficient in Arap1 or Arap2, or with combination deletions of multiple ARAP proteins would clarify whether there is a significant role for the ARAP family of proteins in hematopoiesis.
We find that deletion of Arap3 in stromal and osteoblastic cells using Prx1-Cre does not affect steady-state hematopoiesis or HSC homeostasis. Furthermore, deletion of Arap3 in endothelial cells using VEC-Cre does not impact its ability to support wild-type HSCs. However, high efficiency of Arap3 deletion in f/2;VEC mice results in embryonic lethality, suggesting ARAP3 plays an important role in developing endothelial cells, in agreement with previously published data [29,30]. Since ARAP3 deficiency in embryonic endothelial cells disrupts angiogenic sprouting and vascular structure, this developmental-related dysregulation may indirectly affect HSC emergence, development, and maintenance in the adult mouse. However, our studies do not exclude the possibility that compensatory mechanisms may be upregulated in surviving f/2;VEC mice sometime between HSC emergence and adulthood. One approach to answer this question would be to induce the excision of Arap3 in endothelial cells when ARAP3 is no longer required for vasculature development. Future investigations are warranted to examine the emergence of definitive hematopoietic progenitors and HSCs from the endothelium in the AGM region of Arap3 null embryos and further our understanding of the role of ARAP3 in endothelial cells.
Our results demonstrate that the ARAP3 R302,303A mutation (KI/KI) that disrupts ARAP3 recruitment to the plasma membrane or activation by PI3K markedly impairs HSC function, while loss of ARAP3 does not. This could be due to compensatory mechanisms or changes in gene expression and signaling pathways of rare surviving KI/KI mice. Another possibility for compromised HSCs in KI/KI mice might be that the ARAP3 R302,303A mutant acts in a dominant negative manner to prevent translocation of other interacting players [51][52][53][54][55] to the plasma membrane, which may affect HSC function. For example, ARAP3 can bind the phosphatase SHIP2, a negative regulator of PI3K signaling [53,54,56], as well as CIN85 and Odin, both shown to be involved in receptor endocytosis and motility [51,55,[57][58][59]. Although ARAP3 has not been shown to act as a scaffold for the plasma membrane translocation of any interacting proteins, and the in vivo relevance of these interactions remains to be tested, the fact that we see a cell-intrinsic defect with KI/KI HSCs but not with f/2;Vav HSCs in the BMT assays speaks in favor of this possibility. Lastly, it is possible that ARAP3 plays an important role in a cell type other than hematopoietic, endothelial, stromal, or osteoblastic cells to impact HSC function. One way of determining whether the phenotype of KI/KI HSCs is due to a cell-autonomous effect of the R302,303A mutant in HSCs would be to generate a Vav1-Cre-driven conditional KI/KI mouse to genetically study this question.
RhoA has been characterized in vivo to regulate migration and chemotaxis of mature hematopoietic cells [27,28], as well as HSPC engraftment, multi-lineage repopulation and cell survival [9,14,15]. ARAP3 is not the only GTPaseactivating protein that targets RhoA, and our data suggest it is not a major regulator of RhoA activation in HSPCs. Other GAPs, such as p190B RhoGAP, play important roles in mediating HSPC function through its inactivating activity on RhoA [20,21]. Due to the large number of regulators for RhoA, it is likely that each acts in its own individual temporal-and spatial-specific manner. The possibility of redundancies between the various GAPs also exists [60], such that changing the dynamic by ablating one GAP is not enough to alter the process of normal hematopoiesis. It would be interesting to investigate whether deletion of multiple GAPs in hematopoietic cells would result in greater deficiencies than migration or engraftment alone.
As a dual GTPase-activating protein, ARAP3 targets Arf6 as well as RhoA. Arf6 has mostly been studied in non-hematopoietic cells with regard to its role in membrane trafficking and the cell actin cytoskeleton [61][62][63]. Like RhoA, it is actively involved in cell migration, adhesion, proliferation and cytokinesis [64][65][66]. However, the potential role of Arf6 in HSPCs and hematopoiesis has not been well established. One study showed that the decrease of active Arf6-GTP in platelets is critical to the activation of Rho GTPases that is necessary for cytoskeletal rearrangements preceding full platelet function [67]. It is important to further investigate and understand the role of Arf6 in hematopoiesis, particularly in HSCs.
ARAP3 has been implicated in the regulation and progression of several human diseases, including defense against bacterial infection, diabetes and gastric carcinoma, by capitalizing on the ability of ARAP3 to manipulate vesicle internalization and cell invasion [68][69][70]. Dysregulation of Rho family GTPases and their regulators have also been correlated with human blood disorders and tumorigenesis [71][72][73][74][75]. While aberrant expression of ARAP3 has not yet been found in blood disorders, its ability to regulate the actin cytoskeleton makes it a potential target for the dysregulation of homeostatic cell functions. Thus, continued study of ARAP3 in normal and abnormal hematopoiesis will be important to elucidate a more comprehensive understanding of its role in the blood system.
Generation of Arap3 transgenic mice
Arap3 flox/flox and Arap3 KI/KI mice were generated as described previously [29]. Vav1-Cre mice were originally generated by Dr. Thomas Graf [39] and backcrossed to C57Bl/6J background for 8 generations. VEC-Cre mice were kindly provided by Dr. Nancy Speck [44] and backcrossed to C57Bl/6J background for 8 generations. These two strains of mice were crossed to generate Arap3 flox/+ ;Cre tg mice, which were then crossed to Arap3 flox/flox mice to generate Arap3 flox/flox ;Cre tg conditional knockout mice. Initial studies presented in S1-S3 Figs of f/f;Vav and f/ f;VEC mice were done in mixed Bl6/129 background, while studies in main figures were later performed on a pure C57Bl/6J background following backcrossing for 8 generations. Arap3 flox/+ mice on the pure C57Bl/6J background were crossed with CMV-Cre mice on the C57Bl/6J background (The Jackson Laboratory) to generate Arap3 +/2 mice. These mice were used to generate Arap3 flox/2 ;Vav1-Cre tg mice, Arap3 flox/2 ;VEC-Cre tg and Arap3 flox/2 ;Prx1-Cre tg conditional knockout mice (all on a pure Bl6 background) that will ensure a more complete deletion efficiency. Prx1-Cre mice on a C57Bl/6J background were purchased from The Jackson Laboratory.
The animal studies were carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee of the Children's Hospital of Philadelphia.
For quantitative real-time PCR, total RNA was isolated from FACS-sorted bone marrow cells, CFC assays or hematopoietic tissues using Trizol Reagent (Invitrogen Life Technologies) followed by isolation with the RNeasy Mini kit (Qiagen). cDNAs were produced using BioRad iScript kit and qRT-PCR reactions were performed on an Applied Biosystems 7900HT real-time PCR system using Sybr-Green detection with the following primers:
Complete blood counts
Peripheral blood was collected by retro-orbital bleeding into capillary blood collection tubes with EDTA (BD). CBC analysis was performed using the mouse setting on a HemaVet 950 machine (Drew Scientific).
Cell sorting and flow cytometry
Cells from either peripheral blood or hematopoietic tissues were lysed of red blood cells, then stained with surface markers on ice and washed in PBS with 2% bovine calf serum. Surface markers used to identify cell populations are: CD3e-PE for T-cells, CD19-APC for B-cells, Gr1-PE and Mac1-APC for myeloid cells (eBiosciences), and propidium iodide for viability. This method was also used for peripheral blood analysis of transplanted mice with the addition of CD45.1-PE-Cy7 and CD45.2-FITC antibodies. Flow cytometry was performed on a FACS Canto analyzer (BD).
For cell sorting, bone marrow cells were first depleted of lineage-positive cells using biotinylated lineage cocktail and streptavidin-coupled Dynabeads (Invitrogen) per manufacturer's protocol. Lineage-negative (Lin 2 ) cells were then stained for LSKs as described above and sorted on a FACS Aria Cell Sorter (BD). All analysis of FACS data was performed using FlowJo software (TreeStar).
Colony-forming cell (CFC) assays
Unfractionated bone marrow cells or sorted LSK cells were plated at a concentration of 15,000 or 200 cells per plate, respectively, in duplicate using semisolid methylcellulose (Methocult M3434, StemCell Technologies) containing SCF, IL-3, IL-6 and erythropoietin. Cells were incubated at 37˚C, 5% CO 2 with high humidity, and colonies were enumerated after 10-12 days in culture.
Competitive bone marrow transplantation (BMT), serial BMTs, and reverse BMTs
Bone marrow cells (CD45.2 + ) were harvested and sorted for LSK surface markers as described above. 500-1000 LSK cells were mixed with 350,000 competitor bone marrow cells (CD45.1 + ) and injected into each lethally-irradiated (a split dose of 10Gy, 137 Cs source) [76] F1 recipient mouse (CD45.1 + CD45.2 + ). Donor-derived reconstitution in the periphery was measured by flow cytometry every 4 weeks post-transplant. At 12-16 weeks, recipient mice were sacrificed and analyzed for donor HSPCs, as described above with the addition of CD45.1-PE-Cy7 and CD45.2-FITC antibodies. For secondary transplantation, 2610 6 unfractionated bone marrow cells from pooled primary recipient mice were transplanted into lethally irradiated F1 recipients. Reconstitution was again measured every 4 weeks post-transplant and data of secondary transplant endpoints are shown where mentioned. Tertiary transplants were performed similarly. In reverse BMTs, 1610 6 total bone marrow cells from CD45.1 + wild-type donor SJL mice were transplanted into lethally irradiated f/f and f/2 control or f/f;VEC and f/2;VEC (CD45.2 + ) CKO recipients. Reconstitution and the percentage of CD45.1 + LSK and SLAM LSK cells were measured 8 weeks after transplantation. | 5,919.2 | 2014-12-26T00:00:00.000 | [
"Biology"
] |
An overview of image-based phenotyping as an adaptive 4.0 technology for studying plant abiotic stress: A bibliometric and literature review
Improving the tolerance of crop species to abiotic stresses that limit plant growth and productivity is essential for mitigating the emerging problems of global warming. In this context, imaged data analysis represents an effective method in the 4.0 technology era, where this method has the non-destructive and recursive characterization of plant phenotypic traits as selection criteria. So, the plant breeders are helped in the development of adapted and climate-resilient crop varieties. Although image-based phenotyping has recently resulted in remarkable improvements for identifying the crop status under a range of growing conditions, the topic of its application for assessing the plant behavioral responses to abiotic stressors has not yet been extensively reviewed. For such a purpose, bibliometric analysis is an ideal analytical concept to analyze the evolution and interplay of image-based phenotyping to abiotic stresses by objectively reviewing the literature in light of existing database. Bibliometricy, a bibliometric analysis was applied using a systematic methodology which involved data mining, mining data improvement and analysis, and manuscript construction. The obtained results indicate that there are 554 documents related to image-based phenotyping to abiotic stress until 5 January 2023. All document showed the future development trends of image-based phenotyping will be mainly centered in the United States, European continent and China. The keywords analysis major focus to the application of 4.0 technology and machine learning in plant breeding, especially to create the tolerant variety under abiotic stresses. Drought and saline become an abiotic stress often using image-based phenotyping. Besides that, the rice, wheat and maize as the main commodities in this topic. In conclusion, the present work provides information on resolutive interactions in developing image-based phenotyping to abiotic stress, especially optimizing high-throughput sensors in image-based phenotyping for the future development.
Improving the tolerance of crop species to abiotic stresses that limit plant growth and productivity is essential for mitigating the emerging problems of global warming.In this context, imaged data analysis represents an effective method in the 4.0 technology era, where this method has the nondestructive and recursive characterization of plant phenotypic traits as selection criteria.So, the plant breeders are helped in the development of adapted and climate-resilient crop varieties.Although image-based phenotyping has recently resulted in remarkable improvements for identifying the crop status under a range of growing conditions, the topic of its application for assessing the plant behavioral responses to abiotic stressors has not yet been extensively reviewed.For such a purpose, bibliometric analysis is an ideal analytical concept to analyze the evolution and interplay of image-based phenotyping to abiotic stresses by objectively reviewing the literature in light of existing database.Bibliometricy, a bibliometric analysis was applied using a systematic methodology which involved data mining, mining data improvement and analysis, and manuscript construction.The obtained results indicate that there are 554 documents related to image-based phenotyping to abiotic stress until 5 January 2023.All document showed the future development trends of image-based phenotyping will be mainly centered in the United States, European continent and China.The keywords analysis major focus to the application of 4.0 technology and machine learning in plant breeding, especially to create the tolerant variety under abiotic stresses.Drought and saline become an abiotic stress often using image-based phenotyping.Besides that, the rice, wheat and maize as the main commodities in this topic.In conclusion, the present work provides information on resolutive interactions in developing analysis based on multiple parameters.The final step involved the development of the results.Meanwhile, the literature review only used the related references with the focus keywords that result from bibliometric analysis.The bibliometric methodology applied for complete each step is detailed described in the following sub-sections.
Literature search strategy
The literature is derived from document mining on Scopus.On January 5, 2023, bibliographic data mining was conducted.Scopus centered its data mining efforts on two search categories with the concept of all sources and endless time.The first search field is IBP, and the second search field is abiotic stress or codenamed ALL (image-based AND phenotyping) AND ALL (abiotic AND stress).The range of collecting data is less than 5 January 2023.Combining the two search fields resulted in 554 documents related to IBP to abiotic stresses.The documents obtained from Scopus data mining include 482 journals, 41 books, 20 Book Series, 10 conference proceedings, and one trade journal.
Meanwhile, the data mined includes citations, bibliographical information, abstract, keywords, and other data (trade names and manufacturers, accession numbers and chemicals, conference information, and references).The citations include author name, author Scopus ID, document title, year, source title, volume, issue, page, citation count, source and document type, and publishing stage.The bibliography includes affiliation and publisher.The abstract and keywords include author keywords and index keywords.
Refinement of the search results
Mining data from Scopus was enhanced using OpenRefine 3.6.1 (BSD 2-Clause License) and JAVA.This software will eliminate errors and irrelevant articles from mined data files.The process is focused on the Author and index keywords in the mined data.The data in both categories are split and grouped against words with the same meaning or purpose.After that, the keywords that have been grouped are combined back into one.It also improves the accuracy of Scopus data mining results.Therefore, the notion of mining data refining and filtering is performed semi-manually.
Bibliometric analysis
There are five approaches to bibliometric analysis: general information, global topic development, and topic development based on keywords.The analysis uses three types of software, namely, Tableau, VOSviewer, and Rstudio (version 3.6.1,R Studio Inc., Boston, MA, USA), with the bibliometrix package.Tableau is the widely used visualization software.This software can summarize the advantages of various software such as GIS, FRED, Infographics, or Excel.Consequently, the data analysis sector highly values this program [56].In this study, Tableau only analyzes the number of documents per country.VOSviewer focuses on document data and relational knowledge units for document construction [57].VOSviewer creates scientific knowledge maps that show the relationship among the literature on this topic.This software has large-scale graphical presentation and multifunctional adaptability to source data from different database formats [42](.The application of this software focuses on clusters of interaction and collaboration across countries, interactions among keywords, and citation interactions among publishers.Other analyses use Rstudio and the bibliometrix package.This package can fit the bibliometric concept, and it is free.Furthermore, this software may collect a variety of data mining sources, improve reference disambiguation through string-based algorithms, implement direct and tri-cited analysis, and use a hybrid method that combines bibli-ometric and semantic approaches.The last development includes the detection of term explosions by thematic mapping, which can smooth topics and evolution, and latent semantic analysis [58].
Expansion of the subject of IBP to abiotic
The 554 documents obtained from Scopus containing some generic information.From 2010 to 2023, 251 sources were used to construct this topic's database with 64,097 references, 2765 writers and 1566 author's keywords are identified.In addition, 16 documents are written by a single author.This topic also has 37 % international co-authorship and 6.46 co-authors per document.Fig. 1.Scheme of the bibliometric approach used to review the topic of IBP to abiotic stresses.
M.F. Anshori et al.
Moreover, the average number of citations per paper on this subject is 28.85.
The development of the number of documents and citations which become a fundamental topic in bibliographic analysis is presented in Fig. 2.These two pieces of information become the starting point for analyzing the importance of an issue in the future.The number of documents produced on image-based phenotyping (IBP) focusing abiotic stresses has drastically increased since 2013 (Fig. 2a).The sharpest increase occurred from 2019 to 2021.However, the increase has been sloping from 2021 to 2022.This kind of decrease it is probably related to the global COVID-19 Emergency which has prevented the normal performance of research activities, especially in a field such as that of phenotyping for the analysis of plant responses to abiotic stresses which require continuous data acquisition over time.Meanwhile, the highest number of citations was recorded in 2013 and 2022 (Fig. 2b).Based on document result analysis, IBP for abiotic stress has become crucial in recent years.The number of publications still has an increasing trend with high citation in 2022.It is in line with dynamic climate change [59].So, a systematic and straightforward approach is needed to detect physiological mechanisms and plant tolerance to these stresses, like the IBP concept.According to Costa et al. [45], plant phenotyping has more rapid trends than other approaches in plant evaluation and has a strong linkage to precision agriculture.Therefore, the topic of IBP to abiotic stresses has good momentum in its development, particularly at the end of the citation momentum occurring in 2022.
Development of the topic of IBP to abiotic stresses globally across countries
The global IBP to abiotic stresses can be seen from the number of documents per country, the trend of document development in each country, corresponding issues across countries, and co-responding interactions among countries.In general, the product of papers related to IBP to abiotic stresses is evenly distributed on the European continent.By contrast, the lowest development associated with this topic document is found on the African continent.However, when analyzed by country, the top five countries with the highest number of documents related to the topic are scattered across various continents, apart from the African continent, including the European continent (282 documents, especially in Germany (59 documents), United States (155 documents), China (103 documents), India (90 documents), and Australia (64 documents); Fig. 3 and Supplementary 1).Based on the production of papers per year, the beginning of the development of this topic occurred in 2010.These results follow the reports of Zhang et al. [42], Costa et al. [45], Kolhar and Jagtap [54], who state that the United States and China are the largest countries for the development of precision plant phenotyping.However, in this review, China shows a drastic increase in slope every year, particularly from 2020 to 2022.
Based on interactions among countries, there were five main clusters (Fig. 4).The first cluster is centered in Germany (purple in Figure 4).This country is the first to initiate the development of the topic of IBP to abiotic stresses (Fig. 5).It is also a line with Zhang et al. [42] focused on Agriculture multispectral technology.The concept of IBP is significantly dependent on image sensor technology.It indicated the initial idea could be learning in Germany.Then, the development extends to other European countries as the second cluster (Fig. 5).This cluster is centered in France, which consists of the Netherland, Belgium, Poland, Spain, Italy, and Switzerland (red cluster in Fig. 4).This result also stated by Costa et al. [45] that United Europe (EU) had the most significant publication number on prior 2019.However, after 2019, the United States (US) production increased dramatically.The United States is the focal point of the third cluster (blues cluster in Fig. 4).In general, the US interacts with other countries in its cluster, such as Australia, Canada, Mexico, and South Korea, which were only developed in over 2020 (Fig. 5).Nevertheless, the US article production is significantly more than in other countries except China.China has grown after the US cluster but not ahead of Australia and South Korea.The China cluster consists of Iran, Argentina, Denmark, the Czech Republic, Saudi Arabia, Egypt, Pakistan, and Slovakia (Fig. 5).Based on all country clusters, the topic of IBP to abiotic stress centered in UE in prior 2019 year.Then, this concept rapidly comprehensive to other countries, especially massive population countries like the USA, China (1412.3 million), and India (139.3 million).It also fits with the countries' rising populations (332.3 million; Statista 2023).Consequently, they have a considerable quantity of publications on this subject.
Development of keywords on IBP to abiotic stresses
The results of keyword interaction analysis use an occurrence limit of 15, so there are 72 appropriate keywords.After that, the keywords are filtered and produce 51 that interact with each other.The results of the interaction of these keywords show four color groups: blue, red, green, and yellow (Fig. 6).However, the yellow group is minor and only consists of "Zea mays" and "biomass".
The blue group focuses on phenotype keywords as the main keywords of the blue group and is also the main center in analyzing these keywords.Apart from the phenotype, the dominant keywords in the blue cluster are "genetics", "genotype", "crop", "plant breeding", "phenomics", "wheat", "rice", "quantitative trait locus", and "genomics".All members of the keywords in this group are closely related to the direction of using the IBP method in abiotic stress studies.In general, this concept benefits plant breeding programs, especially in the phenotyping or phenomics of a plant line [ [26,60,61]].A close interaction between plant breeding keywords with phenotype and phenomics also characterizes it.The conventional phenotyping approach to plant breeding needs to improve its accuracy so that IBP can improve its selection accuracy [ [60,61]].In addition, this cluster also explains that phenotype and plant breeding are closely related to genetic concepts, such as genomics and quantitative trait locus (QTL).Plant breeding is closely associated with genetic traits passed down from generation to generation, so the development of phenotyping with precision will involve genetic factors [61][62][63].Several studies have carried out this concept, particularly in developing QTLs and genome-wide association studies on a character [64][65][66].Meanwhile, the keywords rice, and wheat show that the use of the concept of IBP to abiotic stress for breeding purposes is widely practiced in these two commodities.This was also reported by Kim et al. [ [34]].The articles related to the two commodities also follow the great demand for these commodities as the world's staple food [67].
The red clusters focus on physiology, drought, and metabolism.Other members of this cluster consist of "genetic variation", "adaptation", "controlled study", "arabidopsis", "plant roots", "plants", "chlorophyll", and "growth, development, and aging".Based on all keyword members, this group focuses on physiological aspects, types, and stress parameters used in the IBP concept of abiotic stress.For example, giving stress to plants will disrupt plant physiological and metabolic processes so that plants will produce specific symptoms in response to this stress [68][69][70].These response symptoms can be detected through an IBP approach so that several IBP studies will validate the results of this analysis on physiological and metabolic parameters [71][72][73].
The dominant stresses that are often detected with this concept are drought and salinity.The joining of drought and salinity tolerance in this cluster reflects this.Drought and salinity generally have the same response, namely water deficit.Water plays a critical role in plant physiological processes, so a lack of water will inhibit and damage existing metabolic processes [74][75][76][77].There are two visible effects and symptoms when experiencing a water deficit: the chlorophyll content of leaves and the development of plant roots.Water deficit will cause photosynthesis to be hampered, so the photosynthate produced is not balanced with the photosynthesis consumed.This will cause oxidative stress, damaging the leaf chlorophyll, so the leaves experience senescence [78].Low photosynthate produced, and water content will affect cell division so that plant development will be hampered [8,75,[79][80][81].The roots, as the part that interacts directly with water, will show the most development growth due to progressive soil drying [75,82].This underlies the chlorophyll content of leaves and root development as crucial aspects of plant growth in water deficit stress.Therefore, IBP in plant breeding must be validated by observing plant physiology and growth, namely chlorophyll, photosynthesis, plant root, and growth, development, and aging.This is also reflected in the interactions that occur between the concepts of physiology and plant breeding in Fig. 7.
The last large cluster (green) has a lot of keyword interaction centers, namely high-throughput phenotyping, remote sensing, deep learning, machine learning, and image analysis.In addition, other cluster members consist of "agricultural robots", "agriculture", "image segmentation", "hyperspectral imaging", and "plant disease".In general, this cluster is a method of development of IBP.When viewed from many keywords, this group of keywords develops concurrently.The concept of high-throughput phenotyping is a trend carried out in IBP [60,72,[83][84][85][86][87].In general, the HTP is a nondestructive and rapid approach of continuously monitoring and measuring multiple phenotypic traits of multiple plants related to the growth, yield and adaptation to biotic or abiotic stresses.This concept uses sensor digitization combined with automation principles to increase the accuracy of these observations, especially in assessing plant productivity, pigment content, tolerance, and disease [85][86][87].This is the basis for why image segmentation, hyperspectral imaging, and plant disease are included in this cluster.In addition, the concept of high-throughput phenotyping is further enriched by the acquisition, processing and analysis of big data, namely remote sensing, deep learning, and machine learning [84,85] (Jayasinghe et al., 2020; Jangra et al., 2021).Remote sensing the science and technology of obtaining reliable information about physical objects (i.e., crops/plants) and the environment through the acquisition, processing and interpreting of data sensed from space (e.g., satellite-based) or sky (e.g., aircraft-and drone-based) [83][84][85]87].Remote sensing technologies allow the acquisition of time-series data over large areas in a short period of time, making it suitable for field applications.Therefore, using high-throughput phenotyping technology based on big data (machine and deep learning) analysis will make it easier for plant breeders to assemble a variety, including in supporting climate change research (Fig. 7).
These keywords can also be clustered with the concept of multiple correspondence analysis as a conceptual structure analysis (Fig. 8).The result shows two clusters.The keywords "plant", "metabolism", "remote sensing", "quantitative trait loci", "plant breeding", "genomics", "climate change", and "deep learning" are the keywords with the most incredible diversity in the central cluster (colored red, Fig. 8).By contrast, the second cluster only consists of two keywords: "crop" and "crop agricultural".Conceptual structure analysis focuses on the outermost point of the cluster.This point is the highest diversity of the dimensional combinations of a variance partition, so the outer point in this analysis can be a conceptual reference for future research progress [58].Therefore, the results highlight that in future it will be important to apply deep learning methods for analysing remotely sensed data in order to relate the genotypic (e.g., metabolic) and phenotypic (i.e., morpho-physiology) responses of plants to abiotic stresses.This will be mandatory to obtain genetic gains in plant breeding for more resilient varieties or species.
Overview of image-based phenotyping to abiotic stresses
The IBP in the topic of 4.0 adaptive technology is the development of precision phenotyping methods to detect the effect of abiotic stresses on the action and symptoms of plant growth comprehensively.This concept can be assess plant potency gradually under stress in each plant phase, especially with colorimetric and segmentation [28,66,83,[88][89][90].This differs from conventional observations, which only use a few representative characters to achieve a goal [91][92][93].Although in some approaches, observations can be made with hand-handling tools and destructive sampling.Both methods have better precision coverage than conventional morphology.However, using this concept requires time, energy, cost, and many samples [66,72,94,95] In plant breeding, each line represents its potential from a large population, so destructive observation of lines becomes ineffective in assessing their optimal potential [60,66,87,96].Therefore, the plant breeding approach is more directed at the concept of IBP, especially in the topic of stress breeding.
IBP applications are also optimized with High-throughput phenotyping (HTP) technology.HTP relies on computer science, engineering, and big data analysis developments to accelerate plant assessment.Taking pictures is carried out automatically and thoroughly depending on the expected goals, so the observation process becomes efficient with high precision [85,97].This concept is also widely applied in the identification of tolerance, especially in drought stress [33,34,61,[98][99][100] and salinity [101][102][103][104].However, the effectiveness of the HTP concept is also influenced by several factors, namely the environment of the observation area or platform, the sensors used, and the method used for data processing and analysis [84,85,105].
Generally, there are three platform concepts in shooting: artificial in a growth chamber or controlled environment, field groundbased, and field aerial-based [72,84,86].The controlled environment is a common and basic platform used in IBP [106].This platform has a reasonable repetition rate, easy to handle, low bias, and good resolution [72].This advantage allows the modeling and evaluation process to be stable and directed.The use of the concept of a controlled environment in drought stress and salinity has also been reported by Haimansis et al. [28], Laraswati et al. [16], and Sakinah et al. [93] on rice; Wu et al. [107] on corn; and Kim et al. [34] on wheat.However, controlled environment platforms require high costs, and monitoring has a limited scope.In addition, these platforms are useful in a preliminary research phase but then the results must be tested in the field for greater agronomic significance.This is due to the high diversity of environments in the field, so precision models in a controlled environment may only sometimes be appropriately applied in the field.Meanwhile, the platforms in the field are divided into two platforms, namely ground and aerial-based [33,72,84].A ground-based platform has the advantages of flexible deployment and good spatial resolution [34,72,84,108].However, the capturing area still needs to be improved, and it takes a long time to cover the entire planting area.In addition, you could state that several ground-based vehicles can damage the plants during their passage in the inter-row or compact the soil (and therefore compromise the growth of the crop and the experimental results when applied for analysing the abiotic stress responses).In contrast, an aerial-based platform has broad coverage, but its effectiveness depends on the weather condition is an additional limiting factor.Moreover, it would be useful to state that UAV technology is very expensive and requires skilled operators, while most of the satellite images can be downloaded for free from the web [33,84].The use of field-based concepts in stressful environments was also reported by Kim et al. [34] on wheat; Qiu et al. [31] and Wu et al. [32] on maize; and Jiang et al. [109] on rice.Field-based environments require exact multispectral and hyperspectral sensors [34,84].This aims to avoid bias due to sunlight, especially in cloudy conditions, so that the capture result can be analyzed in more detail.
Using sensor types and phenotyping methods also influences the effectiveness of HTP in IBP to abiotic stress studies [33,72,83,85] The use of sensors is identical to the capacity to receive light waves.In general, several types of sensor technology are based on these light waves, namely gamma, X-ray, UV, Visible light, near-infrared, and long-wave infrared [72,110].Visible light is a wavelength that is often and commonly used [60].The concept of visible light uses red, green, and blue channels (RGB), as the basis for determining the color of each pixel.However, the RGB approach is widely sensitive to electromagnetic radiation so this sensor can have a higher bias than other sensors [72,85].Several studies have reported the effectiveness of this sensor for abiotic stresses, especially drought and or salinity, namely Haimansis et al. [28], Laraswati et al. [16], Jiang et al. [109], and Sakinah et al. [93] on rice; Munns et al. [111], Paul et al. [112], and Nehe et al. [113] on wheat; and Zhang et al. [114], Qiu et al. [86], Wu et al. [32] and Dodig et al. [115] on maize.
Another sensor that can be used in visible light is Light detection and ranging (LIDAR).LIDAR is an RGB sensor that can be used in dark conditions with high resolution.Sharpness, accuracy, and good resolution make this tool often used in constructing phenotyping based on three dimensions (3D) or remote sensing concepts.The basis of the LIDAR sensor is a laser fired to see the density of the canopy and aspects of plant growth.However, this tool is relatively expensive, complex, and requires a long time and complicated analysis [33,72,83,116] The use of this sensor on abiotic stress (drought and/or salinity) has been reported by Su et al. [117] on maize and Aneley et al. [118]on potatoes.Gamma, X-ray, and UV wavelengths are shorter than visible light [72,83,85].This makes the energy produced in the three waves huge compared to visible light, especially in gamma rays.This enormous energy allows the light to penetrate parts of the plant so that this light can be used to see the inside of metabolic processes, transport pathways, and ion translocation in plants [119].Positron emission tomography (PET) is a sensor used at gamma wavelengths [72].This technology has been used in the animal and human worlds [85,119].In general, PET can precisely monitor the translocation of nutrients in vivo with high resolution [85,[119][120][121].This is very effective for visualization under salinity and heavy metal stress [72,121].In addition, this sensor is not affected by environmental conditions and produces quantitative and visual data [119,121,122].This facilitates the process of interpretation.The application of this sensor to abiotic stress has been reported by Ariño-Estrada et al. [123] in detecting salinity absorption in green foxtails.However, this sensor analysis requires a large amount of money and has not been widely used for plants, so the development process is still not expansive [85].The X-ray CT sensor is used at X-ray wavelengths [72,83,85].This sensor also views morphology and plant tissue parts illustrated in 3D tomographic images [124][125][126][127].This sensor is widely applied to form root architecture found underground [125].Several applications of this method to abiotic stress have been reported by van Harsselaar et al. [128] on potato drought stress, Schmidt et al. [129] on wheat drought stress, and Kehoe et al. [130] on the detection of barley aerenchyma in submergence stress.Meanwhile, the sensor used at the UV level is Chlorophyll fluorescence (ChF) [34,72,85,131,132].This sensor will produce false colors that can be used to predict photosynthetic potential, and the health of the canopy or plant leaves against abiotic stress [34,133].The false color received by the sensor results from the reflection from leaf chlorophyll, indicating the effectiveness of plant photosynthetic phytochemicals [72,134].The effectiveness of the reflection of chlorophyll can be translated as a parameter of the potential quantum yield of photosystem II (Fv/Fm) [72,133,135,136].However, using this sensor requires adjustments to measurements in dark conditions or are strongly influenced by observational light conditions [34,85].Several abiotic stress studies that have used this tool are Phaseela et al. [136] on drought stress and rice salinity; Larouk et al. [137], Sherstneva et al. [138] and Todorova et al. [139] on wheat drought stress; and Kopsell et al. [140] on maize drought stress.
The last category of sensors is those used at a wavelength greater than visible light.This indicates that the energy released at this wavelength is relatively weak.However, the energy provided at these wavelengths can detect surface temperature and biochemical reflections from the tested image [84].Sensors in this category consist of Near-infrared (NIR)/short-wave infrared (SWIR), hyperspectral imaging, thermal infrared, and magnetic resonance imaging (MRI) [72,84].NIR/SWIR is a transition sensor from visible light [141].This sensor will receive reflected radiation from leaf chlorophyll, including leaf reflections transmitted by the canopy from top to bottom, so that architecture, thickness, and leaf moisture content can be detected in this system [83,141,142].This makes observation non-destructive and easy to do.Several studies have reported the use of this sensor in abiotic stress, namely Phansak et al. [143] and Pabuayon et al. [104] on drought stress and rice salinity; Mokhtari et al. [144], Fan et al. [145], and Danzi et al. [146] on wheat drought; and Casaretto et al. [147] on corn drought.However, this type of sensor is sensitive to wind and cloud cover and requires ground background correction [84].Hyperspectral sensors are characterized by a wide wavelength range, which guarantees a greater level of accuracy than NIR, especially in detecting plant resistance and tolerance [72,[148][149][150][151][152]].For such an example, non-destructive screening of phenotypes within the hyperspectral range allows for quantification of secondary metabolites, including phenolic compounds and flavonoids, involved in plant responses and adaptation to both biotic and abiotic stress conditions [153][154][155].The accuracy of this sensor makes them often used in aerial field-based concepts [84,85,149].However, their weakness, apart from the high cost, is that the interpretation of the results requires more in-depth data analysis [85].Thermal infrared and MRI sensors fall into the long-wave infrared subcategory.Thermal infrared has a more detailed level than NIR [72].In general, the higher the sensor's wavelength, the lower the error from the absorption of infrared radiation [85].This further increases the precision in identifying temperature and water status in the canopy or leaves and stomatal conductivity [156][157][158].However, this concept has areas for improvement regarding its impact on the surrounding environment and data reproducibility, requiring a strict and controlled protocol [85,156,157].Several studies have used thermal imaging sensors for abiotic stress, namely, rice against drought stress [159] and salinity [101,160]; on wheat to drought stress [161][162][163][164] and sodic soils [165].Meanwhile, the MRI sensor moves on radio waves to detect water protons in plant metabolism [72,166].This sensor uses nuclear magnetic resonance signals from several atomic nuclei (1H, 13C, 14 N, and 15 N) to generate phenotyping images [85,166].The results from the sensor are based on 3D images.The high sensitivity of this sensor allows the MRI sensor to detect water content, plant health, plant metabolism, and plant nutrient transport [166,167].However, this sensor also has high costs and low resolution for large cells [85,166].The sensor concepts are briefly shown in Table 1.
The method selected for analyzing the large amount of acquired data is the last factor influencing HTP in IBP.HTP will produce comprehensive data because it is done automatically and on a large population.This will raise many decision-making considerations [72,85,168].Suppose the resulting HTP data needs to be analyzed systematically and in-depth.In that case, the final decision of the analysis will be ambiguous, so the analysis requires precise concepts such as machine and deep learning analysis.Both analyses are M.F.Anshori et al. part of artificial intelligence, which aims to streamline complex populations' automatic and systematic decision-making [72,[168][169][170].The concept of analysis is carried out by creating a pattern and system from big data and integrating it with specific algorithms into an automated system [170][171][172].This is very efficient with computer vision data for plant phenotyping, especially for breeding stress plants that carry out large populations in the straining process [72,169].This effectiveness increases by combining big data from other approaches, such as genomics, transcriptomics, proteomics, and metabolomics.Consequently, the adoption of multiple technical and conceptual capacities pertaining to IBP has the potential to promote novel breeding programs for speeding up genetic improvements of crop plants under different environments [169].
In this view, several international and regional networking initiatives have recently flourished to effectively integrate existing phenotyping facilities, technologies, data, methodologies, services, and resources [173,174].As the world's major plant phenotyping hub, the International Plant Phenotyping Network (IPPN; available online at: https://www.plant-phenotyping.org/)aims to foster interaction between academia, industry, policy and general public stakeholders by favoring synergic training activities which are pivotal for advancements in phenotypic data-driven breeding.Under such domain, large investments for plant phenomics research and cross-disciplinary infrastructures have been made in China, Asia, North America and Europe, including China Plant Phenotyping Network (CPPN), Asia-Pacific Plant Phenotyping Conference (APPP, http://www.appp-con.com),North American Plant Phenotyping Network (NAPPN, https://www.plantphenotyping.org) and European Plant Phenotyping Network 2020 (EPPN2020, https://cordis.europa.eu/project/id/731013).Europe is globally recognized as the leader hub in whole plant phenotyping pipeline, providing shared analytic approaches and interoperable frameworks through excellence scientific community [175].For instance, the European Infrastructure for Multi-Site Plant Phenotyping (EMPHASIS, https://emphasis.plant-phenotyping.eu)aims at developing and implementing pan-European plant phenotyping facilities that can intensify cross-national selection gains of climate-resilient crop plants.Within the EMPHASIS consortium, national networks such as the German Plant Phenotyping Network (DPPN, https://dppn.plantphenotyping-network.de), the French Plant Phenotyping Network (FPPN, https://www.phenome-emphasis.fr), the Italian Plant Phenotyping Network (PHEN-ITALY, http://www.phen-italy.it)and the Austrian Plant Phenotyping Network (APNN, https://appn.at)have enhanced the adoption of IBP approaches for in-depth comprehension of the complex genotype x environment x management (GxExM) interactions underlying plant adaptation to abiotic stresses [176].However, a greater international co-ordination is still indispensable to transfer technological and methodological IBP know-how also to emerging countries on which the 4.0 digitalization phenomics have unique potential for overcoming the global food security challenge under unfavorable environmental scenarios [177].
Conclusion and future perspectives
Based on the results of bibliometric analysis of IBP to abiotic stress, this field is projected to continue to evolve over time.It is based on the fact that the quantity of documents and citations for this issue continues to increase, allowing it to develop sustainably.The center of the development of this topic will remain in two countries, namely, China and the United States.This significant increase in publications of the United States, China, and India to this topic also are based on the increased population of their country.These countries have been categorized as the big three of population in the world (United Nation 2022).The impact of global warming is directly on them, so they must stimulate creating adaptive varieties under abiotic stress.Therefore, the publication about IBP to abiotic stress, particularly in US and China, will increase over time.In addition, several countries such as South Korea, Brazil, and Canada will experience a drastic increase in the relatively large number of articles on this topic, with a new development starting in 2020.These results can also motivate Southeast Asian countries to use this topic as an option in developing publications in the 4.0 era.
The results of bibliometric analysis of keywords show that the development of IBP to abiotic stress is closely related to genetics, genomics, plant breeding, and physiology stress.These words are critical aspects in developing adaptive variety under abiotic stress.Genetics, genomics (i.e., qtl, gwas, and genome editing), and physiology have expensive costs to detect the adaptive variety, so in the future, IBP can help the developing activity in creating a variety based on the relationship at the OMICS level.
Based on abiotic stress kind, drought and saline becomes an abiotic stress often using IBP.Both stresses are closely related to water and oxidative stress.The effect of these stresses can be detected with color contrast and segmentation of plant growth based on IBP.The difference in color contrast and segmentation area is evident in tolerant and sensitive varieties.So, this difference helps to predict the degree of adverse stress of crops.Therefore, the use of IBP is more directed toward drought stress and heat stress.
In future results, IBP to abiotic stress will be related to high-throughput phenotyping (HTP).The HTP is the precision agriculture concept to help the plant breeding program.The effectiveness of HTP is highly dependent on three factors: the platform (i.e., controlled environment and field-based), sensors (i.e., RGB, hyperspectral, thermal, MRI, PET or X-Ray) and distance of the acquiring sensor from the target plant/crop (i.e., proximal and remore sensing), and big data analysis applied.The sensor is crucial as the core of the HTP component.The more precision sensor can develop more parameters in the plant.Meanwhile, using big data analysis will make determining plant tolerance to stress easier based on high-throughput phenotyping (remote sensing and image analysis (2D and 3D)).In general, big data analysis and precision agriculture are the powerful and core approaches in the 4.0 technology era, so the concept of IBP depends on the sensor technology's rhythm and big data analysis development.Therefore, this developmental concept will be integrated into plant breeding; thus, the assembly of varieties becomes more precise.
Data availability statement
Data will be made available on request.M.F.Anshori et al.
Fig. 2 .
Fig. 2. Development of documents related to IBP to abiotic stress studies: (a) the number of documents produced per year and (b) the trend of citations per year.
Fig. 3 .
Fig. 3. Countries that have published the most articles on image-based phenotyping under abiotic stresses.
Fig. 4 .
Fig. 4. Interaction and collaboration clusters across nations concerning image-based phenotyping under abiotic stresses.
Fig. 5 .
Fig. 5. Clusters of international interaction and collaboration on image-based phenotyping under abiotic stresses based on time.
Fig. 7 .
Fig.7.The interaction of "plant breeding" keywords to others keywords.
Table 1
Summary regarding the use of sensors in the concept of high throughput phenotypes. | 8,206.8 | 2023-11-01T00:00:00.000 | [
"Environmental Science",
"Agricultural and Food Sciences",
"Computer Science",
"Biology"
] |
Resonance contributions from $\chi_{c0}$ in the charmless three-body hadronic $B$ meson decays
Within the framework of perturbative QCD factorization, we investigate the nonfactorizable contributions to these factorization-forbidden Quasi-two-body decays $B_{(s)}\rightarrow h\chi_{c0}\rightarrow h\pi^+\pi^-(K^+K^-)$ with $ h=\pi, K$. We compare our predicted branching ratios for the $B_{(s)}\rightarrow K\chi_{c0}\rightarrow K\pi^+\pi^-(K^+K^-)$ decay with available experiment data as well as predictions by other theoretical studies. The branching ratios of these decays are consistent with data and other theoretical predictions. In the Cabibbo-suppressed decays $B_{(s)}\rightarrow h\chi_{c0}\rightarrow h\pi^+\pi^-(K^+K^-)$ with $h=\bar{K}^0,\pi$, however, the values of the branching ratios are the order of $10^{-7}$ and $10^{-8}$. The ratio $R_{\chi_{c0}}$ between the decay $B^+\rightarrow \pi^+\chi_{c0}\rightarrow \pi^+\pi^+\pi^-$ and $B^+\rightarrow K^+\chi_{c0}\rightarrow K^+\pi^+\pi^-$ and the distribution of branching ratios for different decay modes in invariant mass are considered in this work.
Since Belle's measurement [5], numerous theoretical studies have been conducted to investigate the large nonfactorizable contributions, the decay characteristic in B + → K + χ c0 and other relevant decay modes. In the light-cone QCD sum rules approach, the nonfactorizable soft contributions in the B → Kη c , Kχ c0 decays were analyzed in the Ref. [8]. Within the perturbative QCD (PQCD) approach, the nonfactorizable contributions to the B meson decays into charmonia including B 0,+ → K ( * )0,+ χ c0 were calculated in the Refs. [9,10]. In the framework of QCD factorization (QCDF), the exclusive decays including the B → χ c0 K were studied in [11][12][13][14][15][16]. From these studies it was observed that infrared divergences resulting from nonfactorizable vertex corrections could not be eliminated [11,12]. Non-zero gluon mass was then employed to regularize the infrared divergences in vertex corrections [13]. While the authors of [16] found those infrared divergences can be subtracted consistently into the matrix elements of colour-octet operators in the exclusive B to P -wave charmonia decays. In Ref. [15], the B → Kχ c0,2 decays were investigated in QCDF by introducing a non-zero binding energy to regularize the infrared divergence of the vertex part and adopting a model dependent parametrization to remove the logarithmic and linear infrared divergences in the spectator diagrams.
The rescattering effects mediated by intermediate charmed mesons were studied in Refs. [17,18], the authors concluded that such effects could produce a large branching ratio for the decay B + → K + χ c0 .
II. FRAMEWORK
Under the factorization hypothesis, the decay amplitude for B → hχ c0 → hK + K − is given by where the denominator [7] are the pole mass and full width of the resonant state χ c0 , the s is invariant mass square for K + K − pair in the decay final state. L R is the spin of the resonances [27,29]. In the rest frame of the resonant state χ c0 , its daughter K + or K − has the magnitude of the momentum as q = 1 2 s − 4m 2 K , and q 0 in Γ(s) is the value of q at s = m 2 0 . The amplitude A(s) = hχ c0 |H eff |B for the concerned quasi-two-body decays in this work can be found in the Appendix. The mass-dependent coefficient C KK (s) is g χc0K + K − /D BW . We have the coupling constant g χc0K + K − from the relation [54,55] where the Γ χc0→K + K − is the partial width for χ c0 → K + K − . For the process B → hχ c0 → hπ + π − , we need the replacement K → π for the Eqs. (1)-(2) and the relevant parameters. The effective Hamiltonian H eff with the four-fermion operators are the same as in [9].
In the rest frame of the B meson, we choose its momentum p B , the momenta p 3 and p for the bachelor state h and χ c0 , as where x B , x 3 , and z are the corresponding momentum fractions, m B is the mass of B meson. The variable η is defined as η = s/m 2 B , with the invariant mass square s = p 2 . For the B +,0 and B 0 s in this work, we employ the same distribution amplitudes φ B/Bs as in Refs. [36,56]. The wave functions for the bachelor states π and K in this work are written as where m h 0 is the chiral mass, p and z are the momentum and corresponding momentum fraction of π and k. The distribution amplitudes (DAs) φ A (z), φ P (z), φ T (z) can be written as [57][58][59][60] where the Gegenbauer moments are chosen as a π 1 = 0, a K 1 = 0.06, a π,K 2 = 0.25 ± 0.15, a π 4 = −0.015 and the paraments follow The Gegenbauer polynomials are defined as where the variable t = 2z −1. The mass-dependent ππ or KK system, which comes from χ c0 , has the distribution amplitude [9] Φ ππ(KK) = 1 with the twist-2 and twist-3 distribution amplitudes φ v ππ(KK) (z, s) and φ s ππ(KK) (z, s) The timelike form factor F χc0 (z, s) is parametrized with the RBW line shape [61] and can be expressed as follows [62][63][64], where m 0 is the pole mass. The mass-dependent decay width Γ (s) is defined as L R is the spin of the resonances, and L R = 0 for the scalar intermediate state χ c0 .
III. RESULTS
The differential branching ratios (B) for the decay processes B → hπ where τ B is the lifetime of B meson. The q h is the magnitude momentum for the bachelor h in the rest frame of χ c0 : The QCD scale follows Λ [7]. For the shape parameter uncertainty of B (s) meson we use ω B = 0.4 ± 0.04 GeV and ω Bs = 0.5 ± 0.05 GeV, which contributed the largest error for the branching fractions. The second one is from the Gegenbauer moments a h 2 in the bachelor meson DAs. The other two error comes from decay width of the resonance χ c0 and the chiral mass m h 0 of bachelor meson, which have a smaller impact to the uncertainties in our approach. There are further errors which are tiny and can be ignored safely, such as minor and disregarded parameters in the bachelor meson (π/K) distribution amplitudes and Wolfenstein parameters.
Mode
Unit Branching ratios Data[7] B + → K + χc0 → K + π + π − (10 −6 ) 0.81 +0.21 We calculate the branching ratios for the decays of B → hχ c0 → hπ + π − (K + K − ) in Table (I), by using the differential branching ratios in Eq. (11), and the decay amplitudes in the Appendix. Compare our numerical results with current world average values from the PDG [7] and the various theoretical predictions in PQCD, LCSR and QCDF in Table (II), and we do some analyses.
We contrast the various theoretical predictions for the B → Kχ c0 cases of the investigated quasi-two-body and two-body decays. The LCSR calculations mainly focus on B + → K + χ c0 and the prediction value is (1.0 ± 0.6) × 10 −4 [8]. Compared with previous PQCD calculations [9,10], we update the charmonium distribution amplitudes and some of the input parameters in this study. Our predictions are smaller than those of [9] and closer to [10]. The QCDF suffers endpoint divergences caused by spectator amplitudes and infrared divergences resulting from vertex diagrams. The different treatment of these divergences as mentioned in the Introduction in [14][15][16] lead to different numerical results. Both our results in this work and the computations above are in excellent agreement with the available data for B + → K + χ c0 and B 0 → K 0 χ c0 .
For the quasi-two-body processes B + → π + χ c0 → π + π + π − and B + → K + χ c0 → K + π + π − , which have an identical step χ c0 → π + π − , the difference of these two decay modes originated from the bachelor particles pion and kaon. Assuming factorization and flavor-SU (3) symmetry, the ratio R χc0 for the branching fractions of these two processes is With the result in Review of Particle Physics [7], one has R χc0 ≈ 0.036. It still fits expectations from our PQCD anticipated ratio In Fig. 2, we show the distribution of branching ratios for decays modes B + → K + χ c0 → K + K + K − . The mass of χ c0 is visible as a narrow peaks near 3.414 GeV. We find that the central portion of the branching ratios lies in the region around the pole mass of the χ c0 resonance as shown by the distribution of the branching ratios in the ππ invariant mass.
IV. CONCLUSION
We studied the nonfactorizable contributions to these factorization-forbidden quasi-two-body decays B → Kχ c0 → Kππ(KK), B s →K 0 χ c0 →K 0 ππ(KK), and B s → πχ c0 → πππ(KK) in PQCD approach in this work. Our predictions for the branching ratios are summarized in Table I and compared with other theoretical results. The obtained branching ratios of B → Kχ c0 decay are essentially consistent with the current data. For the decay involving π orK in the final state not yet measured, the calculated branching ratios will be further tested by experiments in the near future. By utilizing the flavor-SU (3) symmetry to examine quasi-two-body decays with the same intermediate step, we were able to establish the ratio R χc0 for processes B + → π + χ c0 → π + π + π − and B + → K + χ c0 → K + π + π − . The ratio R χc0 is predicted by PQCD to be 0.049, which is close to the value 0.036 reported in Review of Particle Physics. We also display the distribution of branching ratios for various decay modes in invariant mass, and we discover that the majority of the branching ratios are located in the vicinity of the χ c0 resonance's pole mass.
V. ACKNOWLEDGEMENTS
Many thanks to Wen-Fei Wang, Da-Cheng Yan and Jun Hua for valuable discussions.
Appendix A: Decay amplitudes
The concerned quasi-two-body decay amplitudes are given in the PQCD approach by where G F is the Fermi coupling constant, V , s are the Cabibbo-Kobayashi-Maskawa matrix elements, and c i is Wilson coefficients. The amplitudes appeared in above equations are written as M LL eK(π) = −16 M SP eK(π) = −16 with the r c = m c /m B and r 3 = m h 0 /m B . The evolution factors in above formulas are given by The hard functions h a(b) , the hard scales t a(b) , and factor S ab (t) have their explicit expressions in the Appendix of [66]. | 2,590 | 2023-02-09T00:00:00.000 | [
"Physics"
] |